An online learning peer review score aggregation method based on hypergraph neural network

Abstract

This invention discloses a method for aggregating peer ratings in online learning based on hypergraph neural networks, comprising: constructing a hypergraph of social relationships among students and a corresponding first matrix, a hypergraph of peer ratings among students and a corresponding second matrix, and a hypergraph of student self-ratings and a corresponding third matrix; inputting the first, second, and third matrices into a preset deep hypergraph convolutional model to obtain first, second, and third output features; the deep hypergraph convolutional model includes a convolutional network, a graph attention network, and a residual network; performing feature interaction on the first, second, and third output features to obtain modal interaction features; aggregating the modal interaction features and calculating the predicted score for each student based on the aggregated modal interaction features. This invention can achieve accurate prediction of peer ratings in online learning and can be widely applied in the field of peer rating prediction.

Description

Technical Field

[0001] This invention relates to the field of peer review prediction technology, and in particular to an online learning peer review aggregation method based on hypergraph neural networks. Background Technology

[0002] With the support of computer technology, establishing online learning environments has become the mainstream learning method, enabling teachers and students, and students among themselves, to learn online through discussion and communication. Faced with the massive amount of assignments in online learning environments, peer review has become the primary method of assignment evaluation. However, given the vast amount of peer review data and the complex relationships between peers, how to process student evaluation data to more accurately aggregate students' true performance and determine their actual completion status is a current research challenge. While there is considerable research on peer review aggregation algorithms, and some good results have been achieved, these traditional models use feature vectors as input and output the final aggregated score. However, these methods treat each student in isolation, often neglecting the potential relationships between students involved in peer review. They fail to fully capture the complexity of student data and ignore higher-order relationship information between peers. In reality, peer relationships subtly influence the scoring results. Therefore, capturing these relationships and accurately describing the scoring accuracy is crucial, indicating room for improvement in aggregation models.

[0003] Therefore, the above problems urgently need to be solved. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method for aggregating peer ratings in online learning based on hypergraph neural networks, so as to achieve accurate prediction of peer ratings in online learning.

[0005] One aspect of this invention provides an online peer rating aggregation method based on a hypergraph neural network, comprising:

[0006] Construct a hypergraph of social relationships among students and its corresponding first matrix, a hypergraph of peer review of scores among students and its corresponding second matrix, and a hypergraph of self-evaluation of scores among students and its corresponding third matrix;

[0007] The first matrix, the second matrix, and the third matrix are respectively input into a preset deep hypergraph convolutional model to obtain the first output feature, the second output feature, and the third output feature; the deep hypergraph convolutional model includes a convolutional network, a graph attention network, and a residual network;

[0008] The first output feature, the second output feature, and the third output feature are interacted to obtain the modal interaction feature;

[0009] The modal interaction features are aggregated, and the predicted score for each student is calculated based on the aggregated modal interaction features.

[0010] Optionally, the construction of the hypergraph of social relationships among students and the corresponding first matrix, the hypergraph of peer review among students and the corresponding second matrix, and the hypergraph of self-evaluation of student scores and the corresponding third matrix includes:

[0011] Students are treated as vertices of a hypergraph, and social relationships between students are treated as hyperedges. A social relationship hypergraph is constructed, and a first matrix is ​​constructed based on the social relationship hypergraph.

[0012] Students are treated as vertices of the hypergraph, and peer review between students is treated as hyperedges. A peer review hypergraph is constructed, and a second matrix is ​​constructed based on the peer review hypergraph.

[0013] Students are used as vertices of the hypergraph, and student self-evaluation scores are used as hyperedges to construct a score self-evaluation hypergraph. A third matrix is ​​then constructed based on the score self-evaluation hypergraph.

[0014] Optionally, the step of inputting the first matrix, the second matrix, and the third matrix into a preset deep hypergraph convolutional model to obtain the first output feature, the second output feature, and the third output feature respectively includes:

[0015] The first matrix is ​​input into the deep hypergraph convolutional model and processed sequentially through the convolutional network, graph attention network, and residual network. The first output feature is determined by combining the social relationship weight vector obtained by training the deep hypergraph convolutional model.

[0016] The second matrix is ​​input into the deep hypergraph convolutional model and processed sequentially through the convolutional network, graph attention network, and residual network. The second output feature is determined by combining the score mutual evaluation weight vector obtained by training the deep hypergraph convolutional model.

[0017] The third matrix is ​​input into the deep hypergraph convolutional model and processed sequentially through the convolutional network, graph attention network, and residual network. The third output feature is determined by combining the score self-evaluation weight vector obtained by training the deep hypergraph convolutional model.

[0018] Optionally, the process of determining the first output feature, the second output feature, and the third output feature includes:

[0019] The outputs of the convolutional network, graph attention network, and residual network to the first matrix are multiplied by the corresponding social relationship weight sub-vectors in the social relationship weight vector, and then weighted and averaged to obtain the first output feature.

[0020] The outputs of the convolutional network, graph attention network, and residual network to the second matrix are multiplied by the corresponding score mutual evaluation weight sub-vectors in the score mutual evaluation weight vector, and then weighted and averaged to obtain the second output feature.

[0021] The outputs of the convolutional network, graph attention network, and residual network to the third matrix are multiplied by the corresponding score self-evaluation weight sub-vectors in the score self-evaluation weight vector, and then weighted and averaged to obtain the third output feature.

[0022] Optionally, the step of performing feature interaction on the first output feature, the second output feature, and the third output feature to obtain modal interaction features includes:

[0023] The correlation between each feature in the first output feature, the second output feature, and the third output feature is calculated respectively to obtain the first interaction feature, the second interaction feature, and the third interaction feature.

[0024] The first output feature, the second output feature, and the third output feature are paired to perform feature interaction, resulting in the fourth interaction feature, the fifth interaction feature, and the sixth interaction feature.

[0025] The first output feature, the second output feature, and the third output feature are interacted to obtain the seventh interactive feature.

[0026] Optionally, before aggregating the modal interaction features, the method further includes:

[0027] Convert the matrix dimensions of the first to seventh interactive features to the same dimension.

[0028] Optionally, the aggregation of the modal interaction features includes:

[0029] The first to seventh interaction features are aggregated by calculating a weighted average of the graph attention network.

[0030] Another aspect of this invention provides an online learning peer rating aggregation device based on a hypergraph neural network, comprising:

[0031] The first aggregation unit is used to construct a hypergraph of social relationships among students and its corresponding first matrix, a hypergraph of mutual evaluation of scores among students and its corresponding second matrix, and a hypergraph of self-evaluation of scores among students and its corresponding third matrix.

[0032] The second aggregation unit is used to input the first matrix, the second matrix, and the third matrix into a preset deep hypergraph convolutional model, respectively, to obtain the first output feature, the second output feature, and the third output feature; the deep hypergraph convolutional model includes a convolutional network, a graph attention network, and a residual network;

[0033] The third aggregation unit is used to perform feature interaction between the first output feature, the second output feature and the third output feature to obtain modal interaction features;

[0034] The fourth aggregation unit is used to aggregate the modal interaction features and calculate the predicted score for each student based on the aggregated modal interaction features.

[0035] Another aspect of the present invention provides an electronic device, including a processor and a memory;

[0036] The memory is used to store programs;

[0037] The processor executes the program to implement the above-described method.

[0038] Another aspect of this invention provides a computer-readable storage medium storing a program that is executed by a processor to implement the above-described method.

[0039] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium and execute the computer instructions, causing the computer device to perform the method described above.

[0040] This invention constructs a hypergraph of social relationships among students and its corresponding first matrix, a hypergraph of peer review scores among students and its corresponding second matrix, and a hypergraph of student self-assessment scores and its corresponding third matrix. Hypergraphs have significant advantages in characterizing higher-order relationships among peers. Therefore, this invention proposes to use hypergraphs to characterize higher-order relationships among peers in the task of aggregating peer review scores, providing a new approach to fully mining peer relationship information in peer review tasks. This invention also constructs a deep hypergraph convolutional model based on attention mechanisms and residual structures, and uses the deep hypergraph convolutional model to obtain the outputs of social relationships, peer reviews, and self-assessments, respectively. Then, the three outputs are subjected to feature interaction to obtain the aggregation of all information of each student, and the predicted score of each student is calculated based on the aggregated modal interaction features. Accordingly, this invention can quickly fuse the data of the three parties in the peer review task and accurately calculate the predicted score of each student in this peer review. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is a flowchart illustrating an online learning peer rating aggregation method based on a hypergraph neural network, provided in an embodiment of the present invention.

[0043] Figure 2 An example flowchart of an online learning peer rating aggregation method based on a hypergraph neural network provided in this embodiment of the invention;

[0044] Figure 3 Explanatory diagrams of the variables and their meanings provided for embodiments of the present invention;

[0045] Figure 4 This is an example structural diagram of an online learning peer rating aggregation model based on a hypergraph neural network, provided in an embodiment of the present invention.

[0046] Figure 5 This invention provides a schematic diagram of a hypergraph and its corresponding feature matrix in a peer review process.

[0047] Figure 6 This is a structural example diagram of a deep hypergraph convolution model provided in an embodiment of the present invention;

[0048] Figure 7 This is a schematic diagram illustrating a node feature transformation according to an embodiment of the present invention;

[0049] Figure 8 This is an example diagram illustrating an aggregation and prediction process provided in an embodiment of the present invention;

[0050] Figure 9 This is a structural block diagram of an online learning peer rating aggregation device based on a hypergraph neural network, provided in an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0052] Reference Figure 1 This invention provides a method for online peer rating aggregation based on hypergraph neural networks, specifically including the following steps:

[0053] S100: Construct a hypergraph of social relationships among students and its corresponding first matrix, a hypergraph of peer review of scores among students and its corresponding second matrix, and a hypergraph of self-evaluation of scores among students and its corresponding third matrix.

[0054] Specifically, the process of constructing each hypergraph and matrix can include:

[0055] S1. Construct a social relationship hypergraph by taking students as vertices of the hypergraph and social relationships between students as hyperedges, and construct a first matrix based on the social relationship hypergraph.

[0056] S2. Construct a peer review hypergraph by taking students as vertices of the hypergraph and taking the peer review between students as hyperedges, and construct a second matrix based on the peer review hypergraph.

[0057] S3. Construct a score self-evaluation hypergraph by taking students as vertices of the hypergraph and students' self-evaluation scores as hyperedges, and construct a third matrix based on the score self-evaluation hypergraph.

[0058] S110: Input the first matrix, the second matrix, and the third matrix into a preset deep hypergraph convolutional model to obtain the first output feature, the second output feature, and the third output feature respectively; the deep hypergraph convolutional model includes a convolutional network, a graph attention network, and a residual network.

[0059] Specifically, the process of determining each output feature may include:

[0060] S1. Input the first matrix into the deep hypergraph convolutional model, and process it sequentially through the convolutional network, graph attention network and residual network, and combine it with the social relationship weight vector obtained by training the deep hypergraph convolutional model to determine the first output feature.

[0061] Specifically, the outputs of the convolutional network, graph attention network, and residual network to the first matrix are multiplied by the corresponding social relationship weight sub-vectors in the social relationship weight vector, and then weighted and averaged to obtain the first output feature.

[0062] S2. Input the second matrix into the deep hypergraph convolutional model, and process it sequentially through the convolutional network, graph attention network and residual network, and combine it with the score mutual evaluation weight vector obtained by training the deep hypergraph convolutional model to determine the second output feature.

[0063] Specifically, the outputs of the convolutional network, graph attention network, and residual network to the second matrix are multiplied by the corresponding score mutual evaluation weight sub-vectors in the score mutual evaluation weight vector, and then weighted and averaged to obtain the second output feature.

[0064] S3. Input the third matrix into the deep hypergraph convolutional model, and process it sequentially through the convolutional network, graph attention network and residual network, and combine it with the score self-evaluation weight vector obtained by training the deep hypergraph convolutional model to determine the third output feature.

[0065] Specifically, the outputs of the convolutional network, graph attention network, and residual network to the third matrix are multiplied by the corresponding score self-evaluation weight sub-vectors in the score self-evaluation weight vector, and then weighted and averaged to obtain the third output feature.

[0066] S120: Perform feature interaction on the first output feature, the second output feature, and the third output feature to obtain modal interaction features.

[0067] Specifically, the feature interaction process can include:

[0068] S1. Calculate the correlation between each feature in the first output feature, the second output feature and the third output feature respectively, and obtain the first interaction feature, the second interaction feature and the third interaction feature accordingly.

[0069] S2. Perform feature interaction between each pair of the first output feature, the second output feature, and the third output feature to obtain the fourth interaction feature, the fifth interaction feature, and the sixth interaction feature.

[0070] S3. Perform feature interaction on the first output feature, the second output feature and the third output feature to obtain the seventh interactive feature.

[0071] Furthermore, in order to unify the matrix dimensions of various interactive features, the present invention may also include:

[0072] S4. Convert the matrix dimensions of the first to seventh interactive features to the same dimension.

[0073] S130: Aggregate the modal interaction features and calculate the predicted score for each student based on the aggregated modal interaction features.

[0074] Specifically, the process of aggregating modal interaction features may include:

[0075] The first to seventh interaction features are calculated using a graph attention network to aggregate the modal interaction features.

[0076] To describe the invention in more detail, specific examples will be used to illustrate its practical application.

[0077] Reference Figure 2 This invention provides an example flowchart of an online learning peer rating aggregation method based on a hypergraph neural network. (Refer to...) Figure 3The embodiments of the present invention provide explanatory diagrams of the variables used in the following processes and their meanings. (Refer to...) Figure 4 This invention provides an example structural diagram of an online learning peer rating aggregation model based on a hypergraph neural network.

[0078] Specifically, this embodiment of the invention first constructs a hypergraph based on peer review relationship data. Then, it uses deep hypergraph convolutional models based on attention mechanisms and residual structures to efficiently learn representations of the peer review hypergraph for the feature matrices of social relationships, peer review relationships, and self-review relationships. Subsequently, it performs inner and outer product fusion operations on the output results to obtain an aggregation of all information for each student. The obtained features are then passed through a graph attention network, and by continuously optimizing the training parameters, the output results are made to more accurately approximate the existing real labels. In this way, this embodiment of the invention can quickly fuse the data from the three parties in a peer review task. Finally, after training and outputting the entire framework, the student's final score in this peer review task can be predicted.

[0079] The embodiments of the present invention will now be described in three parts.

[0080] 1. Hypergraph construction for peer review.

[0081] 1.1 Reference Figure 5 This invention provides a schematic diagram of a hypergraph and its corresponding feature matrix in the peer review process.

[0082] 1.2 Specific details.

[0083] 1.2.1 Construct a hypergraph of social relationships.

[0084] In online learning peer review tasks, this embodiment of the invention assumes there are N students, denoted by S, S1, S2, ..., S... N Let N represent the students participating in the online class. These students are required to complete the assigned exercises and assignments, and then conduct peer review and self-evaluation. Since not all students have social relationships with each other, during the peer review process, there is a possibility that students with friendly social relationships may be given a higher score than their true value, while those without friendly social relationships may be given a lower score than their true value. This phenomenon undermines the fairness of the peer review task and negatively impacts the assessment of the students' true understanding. Therefore, in the process of constructing and extracting information from the peer review hypergraph, this embodiment of the invention first considers capturing the social relationships between students.

[0085] Faced with the complex and varied small groups of classmates, whose numbers of friends vary, potentially exceeding the pairwise boundaries, this invention employs a hypergraph representation to fully capture relationships beyond pairwise friendships. This method can quickly capture relationships between groups of three, four, or more friendly classmates. For classmates with friendly friendships, this invention uses hyperedges for connection. This not only covers the pairwise relationships found in traditional simple graph representations, but also... Figure 5 As shown in the diagram, students S5 and S7 connected by e3 are friends. In addition, hypergraphs can also capture friendships among multiple parties. For example, students S2, S3, S6 and S7 in hyperedge e2 are friends. Hypergraphs can more intuitively represent this relationship.

[0086] After the hypergraph is constructed, based on the nodes connected by the hyperedges, this embodiment of the invention constructs a matrix W1 corresponding to social relationships. For example, if e1 connects S1, S5, and S8, meaning these three students are friends, then when constructing the matrix, this embodiment of the invention uses 1 to represent the row containing students S1, S5, and S8, and 0 to represent other students, and so on. In this way, this embodiment of the invention extracts the constructed social relationship matrix W1.

[0087] 1.2.2 Construct a hypergraph for the peer review data.

[0088] In an online learning environment, each student needs to grade the assignments of other students. In a real online classroom environment, different students will grade different numbers of assignments. Some students may not have time to grade too many assignments, resulting in fewer assignments being graded, while other students will grade a relatively large number of assignments. In the face of this situation, if a traditional simple graph with pairing is used to construct the graph, there will be certain limitations. Therefore, this embodiment of the invention also considers constructing a graph for peer review using a hypergraph.

[0089] In the scenario of the task performed in this embodiment of the invention, the full score of the assignment may vary. In order to extract information uniformly, this embodiment of the invention converts all students' scores into scores corresponding to a full score of 10. For different students' scores, their converted decimal scores are used as the weight scores of the matrix. At the same time, when constructing the matrix in this embodiment of the invention, students involved in each peer review are represented by 1, and students not involved are represented by 0, thereby constructing a hypergraph, corresponding to the above. Figure 5 For example:

[0090] Super-edge e n-3Connecting S1, S3, S6, and S8, in the matrix constructed in this embodiment of the invention, based on the original data, during the extraction of the information matrix, in e n-3 In that column, the value of the row corresponding to students S1, S3, S6, and S8 who participated in this evaluation is set to 1. Since no other students were evaluated, the values ​​of the matrices corresponding to the other students are set to 0. The information of the other hyperedges is also extracted in this way. In summary, the peer evaluation matrix W2 is obtained.

[0091] 1.2.3 Construct a hypergraph of self-assessment data.

[0092] To obtain more comprehensive and accurate information and to accurately determine students' true scores from multiple perspectives, students typically self-evaluate their assignments in online classrooms. Therefore, this embodiment of the invention also considers students' self-evaluation scores. Usually, after submitting the assignment and before the peer review task begins, the scoring points and corresponding scoring suggestions are announced to the students. Then, during the peer review phase, students can not only evaluate their peers but also evaluate their own assignments based on the scoring points. In this way, this embodiment of the invention can obtain students' self-evaluation data. Based on the above analysis, this embodiment of the invention can also convert the scores into decimal for analysis and divide students' scores into three categories connected by hyperedges: low self-evaluation score (L≤3), medium score (3<L≤7), and high score (L>7). Students in these three score sets are connected by hyperedges.

[0093] like Figure 5 As shown, e n-2 e represents the superedge with lower performance. n-1 e represents a superedge with average performance. n Let W represent the superedges with high scores. Therefore, it can be seen that S1 and S3 have low self-evaluation scores, S2, S4, S5, and S7 have medium self-evaluation scores, and S6 and S8 have high self-evaluation scores. In summary, this embodiment of the invention yields a self-evaluation matrix W3.

[0094] 2. Deep hypergraph convolution model based on attention mechanism and residual structure.

[0095] 2.1 Reference Figure 6 This invention provides a structural example diagram of a deep hypergraph convolution model.

[0096] A deep hypergraph convolutional model can consist of three parts: a convolutional network, a residual network, and a graph attention network. First, in this embodiment, the feature vector of the constructed hypergraph is input into a multi-layer graph convolutional network. Then, the output of the graph convolutional network is used as the input to the graph attention network. Finally, the residual network is constructed. In this embodiment, the vector output by the graph attention network is represented as h. (1) The initial output vector is represented as h (2) Meanwhile, the output vector obtained after convolution is represented as h. (3) The weights w of the three vectors are obtained through neural network training. (1) w (2) w (3) Then, after weighted averaging, the outputs of the interpersonal relationships, peer evaluation, and self-evaluation networks were obtained as follows: and The following embodiments of the present invention will detail how to use attention mechanisms and residual connections to construct deep hypergraph convolutional models.

[0097] 2.2 Specific details.

[0098] First, we introduce the mathematical representation of the information required in hypergraph learning. Unlike simple graphs, hyperedges in a hypergraph connect two or more nodes. A hypergraph is defined as G = (S, E, w), which contains a set of nodes S and a set of hyperedges E. Each hyperedge is assigned a weighted diagonal matrix w. The hypergraph G can be represented as a |S|×|E| incidence matrix W. For nodes s∈S and e∈E, the node degree d(s) and edge degree d(e) in this embodiment are defined as follows: Equations (1) and (2) are given below.

[0099] d(s)=∑ s∈S ω(e)W(s,e) (1)

[0100] d(e)=∑ e∈E W(s, e) (2)

[0101] The embodiments of the present invention use node degree D e and the degree D s Let represent the diagonal matrices of edge degree and node degree, respectively. Since the prediction of student grades is considered a classification task in this task, the node grade labels should be smooth on the hypergraph structure. This task can be expressed as a regularization framework of formula (3):

[0102] arg min f {R emp (f)+Ω(f)} (3)

[0103] Where Ω(f) is the hypergraph regularization, R emp(f) represents the supervised experience loss, and f is the classification function. The regularized Ω(f) is defined as shown in equation (4):

[0104]

[0105] set up And Δ=I-Θ. Finally, the normalized Ω(f) can be written as in formula (5), where Δ is positive semi-definite, usually called the hypergraph Laplace transform.

[0106]

[0107] Since the Laplace's hypergraph Δ is an n×n positive semi-definite matrix, its eigenvalues ​​Δ = ΓΛΓ can be used. T We obtain the orthogonal eigenvectors Γ = diag(γ1, ..., γ). n Meanwhile, the diagonal matrix λ = diag(λ1, ..., λ) n It contains the corresponding non-negative eigenvalues.

[0108] The following embodiment of the invention uses a peer relationship information framework as an example to analyze how to obtain the output result using a deep hypergraph convolutional model based on attention mechanism and residual structure. The process of analyzing the content framework, such as... Figure 6 As shown.

[0109] When passing through the first convolutional layer, the calculation is processed as follows:

[0110] The input h of the matrix (1) The spectral convolution with filter g is expressed as formula (6):

[0111]

[0112] Where ⊙ is the Hadamard product of elements, g(Λ) = diag(g(λ1), ..., g(λ)). n () is a function of the Fourier coefficients. Due to the scale adaptability of neural networks, the convolution operation can be further simplified to formula (7):

[0113]

[0114] Where θ represents the filter parameters, and A is initialized as an identity matrix, meaning that all hyperedge weights are equal. This applies to a hypergraph signal X∈R with n nodes and C-dimensional features. n×C In this embodiment of the invention, the hyperedge convolution can be expressed as formula (8):

[0115]

[0116] Θ is a parameter that needs to be learned during the training process. An Θ filter is applied to the hypergraph nodes to extract features. After convolution, the output H of the convolution can be obtained in this embodiment of the invention.

[0117] After obtaining the first layer output convolution, multiple layers of convolution iterations are continued. In this embodiment of the invention, the super-edge convolution layers can be represented by the following formula (9):

[0118]

[0119] In the above formula, where H (l) ∈R n×C It is the signal of the l-th layer hypergraph convolution, H 0 =X, where σ represents a non-linear activation function. The convolutional part of deep hypergraph convolutional models is primarily based on spectral convolution on the hypergraph. Therefore, convolutional layers can perform transformations from nodes to edges and back to nodes, and the hypergraph structure allows for better refinement of features. Specifically, first, the initial node features H... (l) Perform learnable filter matrix Θ (l) Processing and feature extraction. Then, based on the hyperedge, collect node features to form the hyperedge feature R. n×C′ It is through This is achieved through multiplication. Finally, the relevant hyperedge features of the output node are aggregated to obtain the features of the output node, with the node degree D. e and the degree D s It plays a normalization role during the computation process. Therefore, deep hypergraph convolutional models based on attention mechanisms and residual structures can effectively extract higher-order correlations on the hypergraph through node-to-edge and then-to-node transformations. A specific transformation diagram is shown below. Figure 7 As shown.

[0120] Figure 7 In this context, N1 represents the number of nodes, and N2 represents the number of hyperedges. Based on the above analysis, it can be inferred that the hypergraph convolution framework of this invention has great feasibility. Furthermore, compared to existing methods, when a hypergraph hyperedge connects only two nodes, the hypergraph is simplified to a simple graph, and Δ coincides with the Laplacian of the simple graph, with an overlap coefficient of half. Compared to existing graph convolution methods, the convolution framework of this invention can naturally simulate higher-order relationships between data and effectively utilize and encode these relationships during feature extraction. Compared to traditional hypergraph methods, the framework of this invention has high convolution computation efficiency and does not require the inverse operation of Laplacian Δ.

[0121] Finally, the convolutional variables are output to the GAT network, and after passing through the residual network, the final output peer relationship calculation result is obtained. Assuming the node data obtained after hypergraph convolution is as follows: Here, N represents the number of nodes, and F is the number of node features. The matrix h is N×F in size, representing the features of all nodes, while R represents the features of only one node, so its size is F×1. The output of the graph attention network is a new set of node feature vectors. We set a weight matrix M to represent the relationship between input and output features, without considering graph structure information. If a self-attention mechanism is implemented for each node, then node S... j For node S i The weights can be expressed as follows (10), assuming the predefined relational function is a:

[0122] e ij =a(Mh i Mh j (10)

[0123] To make attention coefficients easier to calculate and compare, this embodiment of the invention introduces a softmax method for all nodes S. i neighboring node S j After regularization, the attention coefficient α ij It can be expressed as the following formula (11):

[0124]

[0125] Attention mechanism is essentially a single-layer feedforward neural network. After adding the LeakyReLU function to the output layer, the complete attention mechanism is obtained as shown in the following formula (12), where a is the weight matrix connecting the layers in the neural network.

[0126]

[0127] The above calculations yield the attention coefficients between different nodes after regularization, which can be used to predict the output features of each node as shown in formula (13):

[0128]

[0129] The information from the extracted feature matrix is ​​used to calculate the initial output vector. Simultaneously, the output vector obtained after the vector undergoes convolution is... Let the vector output after passing through the final graph attention network be denoted as... The weights w of the three vectors are obtained through neural network training. (i) Then, after weighted averaging, the final output vector is obtained. vector This can be expressed as formula (14):

[0130]

[0131] Using the above method, the embodiments of the present invention can similarly obtain the calculated outputs of the peer evaluation hypergraph and the self-evaluation hypergraph. The embodiments of the present invention count the outputs of the deep hypergraph convolutional models for interpersonal relationships, peer evaluation, and self-evaluation as follows: and

[0132] 3. Peer assessment results aggregation.

[0133] 3.1 Reference Figure 8 The present invention provides an example diagram of the aggregation and prediction process.

[0134] 3.2 Specific details.

[0135] The outputs of the social relationship, peer evaluation, and self-evaluation networks obtained above are respectively and Since this embodiment of the invention can employ a semi-supervised training method, the multimodal dataset is divided into training data and test data, with each dataset containing multiple nodes possessing specific features. The model of this embodiment is trained by reducing the gap between the model output and the true labels of the data. Considering the multimodal nature of supernodes and the fusion of interpersonal relationships and peer review data among different supernodes, the information output by the graph attention network in the final layer is integrated to predict the final student's score. The specific analysis follows three steps:

[0136] The first step is to and and Inner and outer product operations are performed to interact features between modalities. These modal interactions are high-order and implicit, making them difficult to represent. The goal of the internal multimodal interaction network is to capture these interactions and extract modality-specific and cross-modality interaction information. First, for a new embedding vector of a single modality, the Gram matrix is ​​used to calculate the relationships between features, resulting in a single-modality fused vector. For any k vectors y1, y2, ... y in n-dimensional Euclidean space k The matrix formed by the pairwise inner products of the k vectors is the Gram matrix of these k vectors. Obviously, this is a symmetric matrix. The Gram matrix is ​​used to measure the characteristics of each dimension and the relationship between the dimensions. In the multi-scale matrix obtained after the inner product, the diagonal elements provide information about each feature map, and the remaining elements provide information about the relationship between different feature maps. Such a matrix can reflect both the number of features and the degree of closeness between different features. Its main representation is as follows: Formula (15):

[0137]

[0138] For each pair of modes, or between two or more modes, there may be interactions between higher-order modes. For newly embedded vectors of different modes, a second-order tensor representing their information interaction can be obtained through outer product. By calculating the outer product among the three new embedding vectors of different modalities, a third-order tensor representing the information interaction among the three can be obtained. This process continues to yield a finite number of tensors of different orders, which can represent the deep interactions between the new embedding vectors of each modality. Assume H(·) denotes the power set operation and for each subset... This invention can learn a multimodal interaction and a series of related information, i.e., factors, from each S. The complementary information of the supervertex consists of many factors. Let each factor be represented by an F'-dimensional vector. The factors can then be calculated as follows:

[0139] First, if S contains only one element, it means that specific information about a certain modality m∈S is being calculated. Embodiments of this invention can calculate information about this specific modality. The calculation formula is as follows (16):

[0140]

[0141]

[0142]

[0143] in b m ∈R F ′, b m，m ∈R F These are the neural networks g. m (·) and g m，m The parameters of (·) Embedding of MLP networks The Gram matrix represents the interaction information between different modalities. It can be viewed as the mean of specific information of mode m.

[0144] In addition, assuming that S contains elements of multiple modalities, this means that the embedding of cross-modal interaction information between different modalities is being computed. Embodiments of the present invention can calculate cross-modal interaction information. The calculation is as follows (17):

[0145]

[0146]

[0147] in U represents the |S| layer cross-validation of the relevant modalities of the MLP. s and b S These are all parameters that graph neural networks can learn.

[0148] The second step, in this embodiment of the invention, is to design a randomly configured neural network that supports tensor input to achieve the conversion of tensors of different orders to vector representations of the same dimension. Through the above process, deep interactions between various modalities can be obtained. However, what is obtained now are vectors of different orders and dimensions. The graph data obtained in the above stages still faces problems such as large scale and multimodality, and an effective graph neural network model and algorithm are urgently needed to obtain high-quality graphs. Therefore, this embodiment of the invention designs a randomly configured neural network that supports tensor input to convert the vectors of different dimensions obtained in the above process... The vector representation after random configuration of the neural network is as follows: Assuming the existence of three modalities, there are ultimately seven possible representations of the fusion vector, namely:

[0149] The third step involves using a graph attention network to weight and average the relationships between the seven fusion vectors within a node, thereby obtaining the integrated information Y contained in that node. i It is expressed as the following formula (18).

[0150]

[0151] By training the parameters, we obtain the scores of N students in the entire peer task. Since this embodiment of the invention uses a semi-supervised training method, it can utilize the labels L that already have real values. i The result Y output by the model in the embodiment of the present invention i The model is then compared and continuously trained, its loss function reduced, and its parameters optimized until it accurately approximates the actual student performance. This trained peer review framework can then be applied to data where the true performance level is unknown, thereby obtaining the aggregated score results for the student in the peer review task.

[0152] Reference Figure 9 This invention provides an online peer rating aggregation device based on a hypergraph neural network, comprising:

[0153] The first aggregation unit is used to construct a hypergraph of social relationships among students and its corresponding first matrix, a hypergraph of mutual evaluation of scores among students and its corresponding second matrix, and a hypergraph of self-evaluation of scores among students and its corresponding third matrix.

[0154] The second aggregation unit is used to input the first matrix, the second matrix, and the third matrix into a preset deep hypergraph convolutional model, respectively, to obtain the first output feature, the second output feature, and the third output feature; the deep hypergraph convolutional model includes a convolutional network, a graph attention network, and a residual network;

[0155] The third aggregation unit is used to perform feature interaction between the first output feature, the second output feature and the third output feature to obtain modal interaction features;

[0156] The fourth aggregation unit is used to aggregate the modal interaction features and calculate the predicted score for each student based on the aggregated modal interaction features.

[0157] This invention also discloses a computer program product or computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform... Figure 1 The method shown.

[0158] In some alternative embodiments, the functions / operations mentioned in the block diagrams may not occur in the order shown in the operation diagrams. For example, depending on the functions / operations involved, two consecutively shown blocks may actually be executed substantially simultaneously, or the blocks may sometimes be executed in reverse order. Furthermore, the embodiments presented and described in the flowcharts of this invention are provided by way of example to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logic flows presented herein. Alternative embodiments are contemplated in which the order of various operations is altered and sub-operations described as part of a larger operation are executed independently.

[0159] Furthermore, although the invention has been described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the described functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in a separate physical device or software module. It is also understood that a detailed discussion of the actual implementation of each module is unnecessary for understanding the invention. Rather, given the properties, functions, and internal relationships of the various functional modules in the apparatus disclosed herein, the actual implementation of the module will be understood within the scope of conventional skill of an engineer. Therefore, those skilled in the art can implement the invention as set forth in the claims using ordinary techniques without excessive experimentation. It is also understood that the specific concepts disclosed are merely illustrative and not intended to limit the scope of the invention, which is determined by the full scope of the appended claims and their equivalents.

[0160] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0161] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0162] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0163] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0164] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0165] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0166] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

Claims

1. A method for aggregating peer ratings in online learning based on a hypergraph neural network, characterized in that, include: Construct a hypergraph of social relationships among students and its corresponding first matrix, a hypergraph of peer review of scores among students and its corresponding second matrix, and a hypergraph of self-evaluation of scores among students and its corresponding third matrix; The first matrix, the second matrix, and the third matrix are respectively input into a preset deep hypergraph convolutional model to obtain the first output feature, the second output feature, and the third output feature; the deep hypergraph convolutional model includes a convolutional network, a graph attention network, and a residual network; The first output feature, the second output feature, and the third output feature are interacted to obtain the modal interaction feature; The modal interaction features are aggregated, and the predicted score for each student is calculated based on the aggregated modal interaction features.

2. The method for online learning peer rating aggregation based on hypergraph neural networks according to claim 1, characterized in that, The construction of the hypergraph of social relationships among students and the corresponding first matrix, the hypergraph of peer review among students and the corresponding second matrix, and the hypergraph of self-evaluation of student scores and the corresponding third matrix includes: Students are treated as vertices of a hypergraph, and social relationships between students are treated as hyperedges. A social relationship hypergraph is constructed, and a first matrix is ​​constructed based on the social relationship hypergraph. Students are treated as vertices of the hypergraph, and peer review between students is treated as hyperedges. A peer review hypergraph is constructed, and a second matrix is ​​constructed based on the peer review hypergraph. Students are used as vertices of the hypergraph, and student self-evaluation scores are used as hyperedges to construct a score self-evaluation hypergraph. A third matrix is ​​then constructed based on the score self-evaluation hypergraph.

3. The method for online learning peer rating aggregation based on hypergraph neural networks according to claim 1, characterized in that, The step of inputting the first matrix, the second matrix, and the third matrix into a preset deep hypergraph convolutional model to obtain the first output feature, the second output feature, and the third output feature respectively includes: The first matrix is ​​input into the deep hypergraph convolutional model and processed sequentially through the convolutional network, graph attention network, and residual network. The first output feature is determined by combining the social relationship weight vector obtained by training the deep hypergraph convolutional model. The second matrix is ​​input into the deep hypergraph convolutional model and processed sequentially through the convolutional network, graph attention network, and residual network. The second output feature is determined by combining the score mutual evaluation weight vector obtained by training the deep hypergraph convolutional model. The third matrix is ​​input into the deep hypergraph convolutional model and processed sequentially through the convolutional network, graph attention network, and residual network. The third output feature is determined by combining the score self-evaluation weight vector obtained by training the deep hypergraph convolutional model.

4. The method for online learning peer rating aggregation based on hypergraph neural networks according to claim 3, characterized in that, The process of determining the first output feature, the second output feature, and the third output feature includes: The outputs of the convolutional network, graph attention network, and residual network to the first matrix are multiplied by the corresponding social relationship weight sub-vectors in the social relationship weight vector, and then weighted and averaged to obtain the first output feature. The outputs of the convolutional network, graph attention network, and residual network to the second matrix are multiplied by the corresponding score mutual evaluation weight sub-vectors in the score mutual evaluation weight vector, and then weighted and averaged to obtain the second output feature. The outputs of the convolutional network, graph attention network, and residual network to the third matrix are multiplied by the corresponding score self-evaluation weight sub-vectors in the score self-evaluation weight vector, and then weighted and averaged to obtain the third output feature.

5. The method for online learning peer rating aggregation based on hypergraph neural networks according to claim 1, characterized in that, The step of performing feature interaction on the first output feature, the second output feature, and the third output feature to obtain modal interaction features includes: The correlation between each feature in the first output feature, the second output feature, and the third output feature is calculated respectively to obtain the first interaction feature, the second interaction feature, and the third interaction feature. The first output feature, the second output feature, and the third output feature are paired to perform feature interaction, resulting in the fourth interaction feature, the fifth interaction feature, and the sixth interaction feature. The first output feature, the second output feature, and the third output feature are interacted to obtain the seventh interactive feature.

6. The method for online learning peer rating aggregation based on hypergraph neural networks according to claim 5, characterized in that, Prior to aggregating the modal interaction features, the method further includes: Convert the matrix dimensions of the first to seventh interactive features to the same dimension.

7. The method for online learning peer rating aggregation based on hypergraph neural networks according to claim 6, characterized in that, The aggregation of the modal interaction features includes: The first to seventh interaction features are aggregated by calculating a weighted average of the graph attention network.

8. An online learning peer rating aggregation device based on a hypergraph neural network, characterized in that, include: The first aggregation unit is used to construct a hypergraph of social relationships among students and its corresponding first matrix, a hypergraph of mutual evaluation of scores among students and its corresponding second matrix, and a hypergraph of self-evaluation of scores among students and its corresponding third matrix. The second aggregation unit is used to input the first matrix, the second matrix, and the third matrix into a preset deep hypergraph convolutional model, respectively, to obtain the first output feature, the second output feature, and the third output feature; the deep hypergraph convolutional model includes a convolutional network, a graph attention network, and a residual network; The third aggregation unit is used to perform feature interaction between the first output feature, the second output feature and the third output feature to obtain modal interaction features; The fourth aggregation unit is used to aggregate the modal interaction features and calculate the predicted score for each student based on the aggregated modal interaction features.

9. An electronic device, characterized in that, Including the processor and memory; The memory is used to store programs; The processor executes the program to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The storage medium stores a program that is executed by a processor to implement the method as described in any one of claims 1 to 7.

Images