Hypergraph-based abnormal grouping detection method and system for computer-supported collaborative learning

Abstract

The present invention discloses a method and system for detecting abnormal grouping of computer-supported collaborative learning based on a hypergraph, and belongs to the field of graph anomaly detection. The present invention inputs the student online classroom interaction data to be detected into a trained abnormal grouping detection model to obtain the abnormal detection results of CSCL groups in the classroom. The present invention realizes the detection of timing anomalies of grouping and interaction in a coding and decoding manner by adding a timing module to the model; by linearly fusing the encoding result of the timing module into the initial representation of the hypergraph module, the hypergraph module can consider fine-grained timing information; by using a hypergraph joint neural network, the encoding of high-order relations is realized and the representation of hyperedges and hypernodes is learned; and by using a timing self-attention mechanism, the timing information of multiple scales is comprehensively considered in the hypergraph module. In addition, the present invention also uses a self-attention mechanism with Token, which can adaptively aggregate the group representations of different numbers of members.

Description

Technical Field

[0001] The present invention belongs to the field of graph anomaly detection, and in particular relates to a hypergraph-based computer supported collaborative learning anomaly grouping detection method and system. Background Art

[0002] Computer Supported Collaborative Learning (CSCL) is considered to be a modern teaching model that improves students' learning efficiency. In recent years, with the rapid development of Internet technology and artificial intelligence technology, exploring how to use computer technology to support collaborative learning has become an important research field that integrates information science and learning science. In CSCL, learners not only learn as individuals, but also emphasize learning as members of a multi-person group. The interaction between learners can be face-to-face or remotely conducted with the help of computers. However, due to the large number of large-scale learning scenarios in online learning, the large number of students and the complex interactions between students, it is easy for teachers to have an insufficient understanding of the relationship between students, which brings challenges to teachers to group students reasonably. At the same time, when the scale of online teaching is too large, it is difficult for teachers to monitor the collaborative learning of each group in real time, which makes it impossible for teachers to adjust and guide the problem group in a timely and accurate manner. Therefore, achieving high-quality grouping in CSCL and timely discovering problem grouping have become an important issue that CSCL needs to solve.

[0003] In order to solve this problem, there is an urgent need for a timely and efficient CSCL abnormal grouping detection system to quickly and accurately find problematic groupings and provide reasonable explanations for the causes of grouping problems. To achieve high-quality grouping, it is necessary to consider many factors, such as the attributes of each student, the interaction between students, and the time each group member participates in group discussions. Therefore, one of the core of building a CSCL abnormal grouping detection system is to consider the various information contained in the entire online class, so as to judge the pros and cons of the grouping situation, help teachers to make correct adjustments or guidance to the problematic grouping in a timely manner, and achieve better teaching results.

[0004] If the students in the whole class and the interactions between them are represented by a graph structure, each collaborative group is a subgraph in the graph structure. In this way, CSCL anomaly group detection can be regarded as an abnormal subgraph detection task. The current abnormal subgraph mainly detects subgraph-level anomalies based on the node and edge information collected in a specific task scenario. Subgraph-level anomaly detection is more challenging than node-level and edge-level anomaly detection. Since the nodes and edges in the abnormal subgraph may be normal, it is necessary to consider the anomaly from the perspective of the entire subgraph.

[0005] In this research context, there are two main directions for anomaly detection research: one is based on shallow learning methods, such as SADE, a two-stage abnormal subgraph detection framework, which combines community discovery with random walk algorithms to learn subgraph representations, and then uses the isolation forest method to detect anomalies; the other is based on deep learning methods, such as HO-GAT, which builds new dependency views to learn the dependencies between nodes and subgraphs, and uses graph attention networks (GAT) to learn the potential representations of nodes and subgraph instances. Compared with deep learning-based methods, shallow learning methods mainly focus on defining anomalies, quantifying graph metrics, and developing methods to capture anomalies based on these metrics. Although simple, they cannot model the complex interdependencies of graphs.

[0006] The current shortcomings of deep learning methods are: 1) They do not take into account the impact of dynamic time information. For example, the time when grouping starts in classroom teaching and the time when group members discuss in groups are important information worthy of consideration; 2) While considering the information contained in the nodes, the rich information contained in the edges is rarely considered. For example, when measuring whether a grouping is reasonable, the interactive information between students in the group is a key factor that cannot be ignored; 3) Lack of explainability, which will result in the teacher being unable to accurately locate the problem when discovering abnormal grouping and promptly adopt corresponding strategies to adjust and guide the grouping. Summary of the invention

[0007] The purpose of the present invention is to solve the problems existing in the prior art and provide a method and system for detecting abnormal grouping in computer-supported collaborative learning based on a hypergraph. When detecting abnormal grouping of CSCL, the method based on deep learning, especially the graph neural network based on graph structure data, is inspired by the fact that it can effectively process graph structure data, capture local and global information, has good generalization ability and other characteristics, and has good performance in abnormal detection tasks. The present invention is further modeled as a special abnormal subgraph detection task, namely, an abnormal motif detection task, and proposes a method for detecting abnormal grouping in computer-supported collaborative learning based on a hypergraph to solve the above-mentioned shortcomings.

[0008] In order to achieve the above-mentioned invention object, the present invention specifically adopts the following technical solutions:

[0009] In a first aspect, the present invention provides a method for detecting abnormal grouping in computer-supported collaborative learning based on a hypergraph, comprising the following steps:

[0010] S1. Obtain the online classroom interaction data of students to be tested, represent students as original nodes, and use the interactions between students as original edges. When three or more original nodes are fully connected, it is considered that there is dense interaction between students. Students with dense interactions form a CSCL group. All interactions are divided into several snapshots according to the time sequence, and all snapshots form a discrete original dynamic graph structure; each snapshot consists of its own student set, interaction set, interaction timestamp list and student feature matrix;

[0011] S2. Take the original nodes in S1 as motif-enhanced nodes, keep the original edges, and take each CSCL group as a virtual hyperedge. After each virtual hyperedge connects the specified motif-enhanced nodes, a motif-enhanced graph view is obtained. The motif-enhanced graph view is transformed into a dual hypergraph to obtain a motif-enhanced dual hypergraph view.

[0012] S3. Obtain the temporal features of all supernodes in the motif-enhanced dual hypergraph view from the student online classroom interaction data to be detected, take the supernode obtained by the original edge as the ordinary supernode, and take the supernode obtained by the virtual hyperedge as the motif supernode, respectively stack the temporal feature vectors of the ordinary supernode and the temporal feature vectors of the motif supernode into a matrix form, and obtain the temporal feature matrix of the ordinary supernode and the temporal feature matrix of the motif supernode, and concatenate the two temporal feature matrices to obtain a complete temporal feature matrix;

[0013] S4. Input the motif-enhanced dual hypergraph view and the complete temporal feature matrix into the trained anomaly grouping detection model, and output the total anomaly score of the students and the total anomaly score of the CSCL groups or interactions. The total anomaly score of the CSCL groups or interactions includes the CSCL group anomaly score and the anomaly score of the interaction between students.

[0014] Based on the above scheme, each step can be implemented in the following preferred specific manner.

[0015] As a preferred embodiment of the above-mentioned first aspect, in the abnormal group detection model, the encoder of the timing module encodes the complete timing feature matrix, and the decoder of the timing module decodes the encoding result of the encoder, and the encoding result of the decoder is used as the timing reconstruction matrix of each snapshot, and the hyperedge features in the motif enhanced dual hypergraph view are used as the student's own features, and the encoding result of the encoder and the randomly initialized supernode features are weighted and summed as the supernode features of each snapshot, and the student's own features with random values ​​and the supernode features of each snapshot are input into the hypergraph joint neural network of the hypergraph module, and the hypergraph joint neural network is composed of a hyperedge convolutional network The hyperedge convolution network and the supernode convolution network have the same number of layers, the output of each layer of the hyperedge convolution network and the output of the corresponding layer of the supernode convolution network are used to calculate the output of the next layer of the hyperedge convolution network, the output of each layer of the supernode convolution network and the output of the corresponding layer of the hyperedge convolution network are used to calculate the output of the next layer of the supernode convolution network, the supernode features of the last layer of the supernode convolution network are used as the encoding of the interactive information by the hypergraph joint neural network, the hyperedge features of the last layer of the hyperedge convolution network are used as the encoding of the student information by the hypergraph joint neural network, and the encoding of the student information by the hypergraph joint neural network is obtained. Get the current snapshot information to be processed, process the historical snapshot information with the timestamp before the current snapshot information into high-frequency scale and low-frequency scale, perform multi-head self-attention mechanism operation on the high-frequency scale and low-frequency scale respectively, concatenate the high-frequency result and low-frequency result obtained by the multi-head self-attention mechanism operation as the student information encoding containing time information, calculate the student information encoding containing time information and the super-node features output by the last layer of the super-node convolutional network through the self-attention mechanism with Token for each CSCL group, and obtain the encoding features of each CSCL group in each snapshot, and calculate the encoding features of all CSCL groups in each snapshot. The supernode features output by the last layer of the supernode convolutional network are concatenated to obtain the interactive information coding with group information coding. The reconstructed hypergraph association matrix of each snapshot is calculated based on the interactive information coding with group information coding and the student information coding containing time information. The temporal anomaly score is calculated based on the complete temporal feature matrix of each snapshot and the temporal reconstruction matrix. The hybrid anomaly score is calculated based on the association matrix of the motif-enhanced dual hypergraph view and the reconstructed hypergraph association matrix of each snapshot. The total anomaly score of the CSCL group or interaction is calculated by weighted summing the temporal anomaly score and the hybrid anomaly score. The latter in the total anomaly score of the CSCL group or interaction is added. The item is used as the CSCL group anomaly score, and the remaining items are used as the student-to-student interaction anomaly scores. The total anomaly score of the student is calculated based on the association matrix of the motif-enhanced dual hypergraph view and the reconstructed hypergraph association matrix of each snapshot; where, is the number of CSCL groups.

[0016] As a preferred embodiment of the first aspect, the super node convolutional network Output of the layer The specific expressions are as follows:

[0017]

[0018] in, is the activation function; Degree matrices of supernodes in the dual hypergraph view are enhanced for motifs; An incidence matrix for motif-enhanced dual hypergraph views; is the hyperedge weight matrix; Degree matrix of hyperedges in dual hypergraph view for motif enhancement; is the first compromise parameter; is the super node convolutional network The output of the layer; Represents the hyperedge convolutional network The output of the layer; and All are super node convolutional networks. The trainable weight parameters of the layer; Indicates transpose.

[0019] As a preferred embodiment of the first aspect, the hyperedge convolutional network Output of the layer The specific expressions are as follows:

[0020]

[0021] in, is the supernode weight matrix; is the second compromise parameter; and All are hyperedge convolutional networks. Trainable weight parameters of the layer.

[0022] As a preferred embodiment of the above-mentioned first aspect, the high-frequency scale and the low-frequency scale are obtained as follows: perform temporal weight decay on historical snapshot information with a timestamp before the current snapshot information, obtain the high-frequency scale after the temporal weight decay, and obtain the low-frequency scale after pooling the high-frequency scale.

[0023] As a preferred embodiment of the first aspect, the first The specific calculation method of the encoding features of the CSCL group is as follows: the token used to save the feature aggregation result Encoding student information including time information Middle The hyperedge features of the CSCL groups and the supernode features output by the last layer of the supernode convolutional network Middle The features of each supernode of the CSCL group are concatenated and calculated through the self-attention mechanism to obtain the output feature. The 0th element of the output feature is used as the first element in each snapshot. Coding characteristics of the CSCL groups.

[0024] Furthermore, the specific calculation method for obtaining the coding features of each CSCL group in each snapshot is as follows:

[0025] ;

[0026] in, In each snapshot, Coding characteristics of the CSCL groups; Represents the calculation of the self-attention mechanism; Represents the token used to save feature aggregation results; Represents the 0th element of the output feature; Respectively represent the student information code containing time information Middle CSCL group Hyperedge features; Respectively represent the super node features output by the last layer of the super node convolutional network Middle CSCL group Supernode features; Represents the number of members in each CSCL group.

[0027] As a preferred embodiment of the above-mentioned first aspect, in step S2, the timing reconstruction matrix of each CSCL group or interaction in each snapshot is subtracted from the corresponding complete timing feature matrix to obtain a first subtraction result, and the L2 norm of the first subtraction result is used as the timing anomaly score of each CSCL group or interaction; the reconstructed hypergraph association matrix of each CSCL group or interaction in each snapshot is subtracted from the association matrix of the corresponding motif-enhanced dual hypergraph view to obtain a second subtraction result, and the L2 norm of the second subtraction result is used as the mixed anomaly score of each CSCL group or interaction; the reconstructed hypergraph association matrix of each student in each snapshot is subtracted from the association matrix of the corresponding motif-enhanced dual hypergraph view to obtain a third subtraction result, and the L2 norm of the third subtraction result is used as the total anomaly score of each student.

[0028] As a preferred embodiment of the above-mentioned first aspect, during the training process of the anomaly grouping detection model, the temporal anomaly score is used as the reconstruction loss of the complete temporal feature matrix, the mixed anomaly score is used as the reconstruction loss of the association matrix of the motif-enhanced dual hypergraph view, the weighted sum of the reconstruction loss of the complete temporal feature matrix and the reconstruction loss of the association matrix of the motif-enhanced dual hypergraph view is used as the total loss of the anomaly grouping detection model, and the parameters of the anomaly grouping detection model are updated based on minimizing the total loss until the total loss converges to obtain a trained anomaly grouping detection model.

[0029] As a preferred embodiment of the above-mentioned first aspect, in step S2, the specific process of constructing the motif-enhanced graph view is as follows: a set of CSCL groups is obtained by searching in all snapshots, and each CSCL group in the CSCL group set is used as a virtual hyperedge, and each virtual hyperedge connects all the original nodes in the corresponding group and the first-order neighbors of all the original nodes in the corresponding group.

[0030] In a second aspect, the present invention provides a hypergraph-based computer-supported collaborative learning anomaly grouping detection system, comprising:

[0031] The data acquisition and processing module is used to obtain the online classroom interaction data of students to be tested, and represents students as original nodes and the interactions between students as original edges. When three or more original nodes are fully connected, it is considered that there is dense interaction between students. Students with dense interaction form a CSCL group. All interactions are divided into several snapshots according to the time sequence, and all snapshots form a discrete original dynamic graph structure; each snapshot consists of its own student set, interaction set, interaction timestamp list and student feature matrix;

[0032] A motif-enhanced dual hypergraph view acquisition module is used to use the original nodes in the data acquisition and processing module as motif-enhanced nodes, retain the original edges, and use each CSCL group as a virtual hyperedge. After each virtual hyperedge connects the specified motif-enhanced node, a motif-enhanced graph view is obtained. The motif-enhanced graph view is subjected to a dual hypergraph transformation to obtain a motif-enhanced dual hypergraph view.

[0033] A complete temporal feature matrix acquisition module is used to obtain the temporal features of all supernodes in the motif-enhanced dual hypergraph view from the student online classroom interaction data to be detected, and the supernodes obtained from the original edges are used as ordinary supernodes, and the supernodes obtained from the virtual hyperedges are used as motif supernodes. The temporal feature vectors of the ordinary supernodes and the temporal feature vectors of the motif supernodes are stacked into matrix form respectively, and the temporal feature matrices of the ordinary supernodes and the temporal feature matrices of the motif supernodes are obtained accordingly. The two temporal feature matrices are spliced ​​to obtain a complete temporal feature matrix;

[0034] The result acquisition module is used to input the motif-enhanced dual hypergraph view and the complete temporal feature matrix into the trained anomaly grouping detection model, and output the total anomaly score of the students and the total anomaly score of the CSCL group or interaction. The total anomaly score of the CSCL group or interaction includes the CSCL group anomaly score and the anomaly score of the interaction between students.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] When using traditional neural network models to detect CSCL anomaly groups, it is impossible to take into account the rich interactive information contained in the application scenario, such as the interaction between students in the online classroom, and it is impossible to make full use of the classroom information, resulting in low accuracy of anomaly group detection. The present invention models the online classroom CSCL scenario as an anomaly motif detection task based on a hypergraph network, uses a hypergraph combined with a neural network to detect anomaly groups in CSCL, and obtains an anomaly score for each CSCL group. The main innovations are:

[0037] 1) Construct the online classroom student collaboration data into a hypergraph. Each group of a specified size in the original dynamic graph is abstracted as a virtual hyperedge for message passing considering group information. Considering that a group is a higher-order unit relative to individual students, the present invention performs a dual hypergraph transformation, so that the encoder can encode higher-order relationships between students, students, and groups, and groups.

[0038] 2) A hypergraph joint neural network is proposed to encode the base-order enhanced hypergraph view composed of students' online classroom interaction data, and collaboratively learn the representation of hyperedges and hypernodes. That is, when learning the representation of students, the interactions related to the students are taken into account; when learning the representation of interactions between students, the students related to the interactions are taken into account, so as to make full use of classroom information;

[0039] 3) Considering the various information contained in the CSCL group, the members and interaction features of each CSCL group are aggregated using a self-attention mechanism with tokens. The output representation of the calculated tokens is used as the representation of the CSCL group and fed into downstream tasks.

[0040] 4) The main function of the present invention is to detect grouping anomalies. The reconstruction loss of the hypergraph association matrix can calculate the reconstruction error of each CSCL group while calculating the reconstruction error of each student and each interaction, which is used as the anomaly score. While obtaining the anomaly score of the CSCL group, the anomaly score related to the group members can be obtained as an auxiliary to provide a certain degree of explainability.

[0041] When using traditional neural network models for CSCL anomaly grouping detection, it is impossible to take into account the rich time information contained in the application scenario, such as the time of class and the time when group members communicate. This will lead to the inability to timely discover problem groups during the class and ignore the anomalies caused by time factors. The present invention divides the online classroom student collaboration data into multiple snapshots according to the time sequence of the class, and proposes coarse and fine-grained timing modules to consider time information. The main innovations are:

[0042] 1) It is proposed to represent the classroom as discrete snapshots to capture dynamic information. The original dynamic graph structure in discrete form tends to model systems where the relationship between entities changes over time in a discrete manner. In this graph, the structure of the graph evolves at different time steps or snapshots to capture the dynamic nature of the network;

[0043] 2) Propose a coarse-grained temporal self-attention mechanism. Each snapshot represents the classroom information of different time periods in the online class. Considering that the students in the same class are unchanged, each student will appear repeatedly in different snapshots, and the same student may contain different information in different snapshots. The present invention adopts a dual-channel self-attention mechanism to capture coarse-grained information at different scales, that is, when considering the current information of each student, it also refers to the historical information;

[0044] 3) Propose a fine-grained timing module. The timing information of the interaction and CSCL group is represented as a matrix, and the encoder and decoder are used to obtain the abnormality of a student or CSCL group at the timing level within the snapshot. For example, a student joins the CSCL group discussion late, which affects the overall learning efficiency of the CSCL group.

[0045] 4) Fuse the fine-grained temporal encoding results with the original information of the class to consider more information when delivering messages. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of the steps of the present invention;

[0047] Figure 2 It is a structural schematic diagram of the abnormal group detection model of the present invention;

[0048] Figure 3 It is a system block diagram of the present invention. DETAILED DESCRIPTION

[0049] In order to make the above-mentioned purpose, features and advantages of the present invention more obvious and easy to understand, the specific implementation mode of the present invention is described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. The technical features in each embodiment of the present invention can be combined accordingly without conflicting with each other.

[0050] In the description of the present invention, it should be understood that the terms "first" and "second" are only used for the purpose of distinguishing descriptions, and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features.

[0051] like Figure 1 As shown, in a preferred implementation of the present invention, the above-mentioned hypergraph-based computer-supported collaborative learning abnormal grouping detection method includes the following steps S1 to S4. The specific implementation process is described in detail below.

[0052] S1. Obtain the online classroom interaction data of students to be tested, represent students as original nodes, and the interactions between students as original edges. When three or more original nodes are fully connected, it is considered that there is dense interaction between students. Students with dense interactions form a CSCL group. All interactions are divided into several snapshots according to the time sequence, and all snapshots constitute the original dynamic graph structure in discrete form; each snapshot consists of its own student set, interaction set, interaction timestamp list and student feature matrix.

[0053] It should be noted that step S1 of the present invention is the processing of student online classroom interaction data, which processes the student online classroom interaction data into a dynamic graph structure. In the embodiment of the present invention, based on the obtained student online classroom interaction data, students are represented as nodes and the interactions between students are regarded as edges. When three or more nodes are fully connected, it is considered that there is dense interaction between students, and the students with dense interaction form a CSCL group. All interactions are divided into snapshots, all of which constitute the original dynamic graph structure in discrete form , among which, Snapshots By Snapshot Student Collection , No. Snapshot Interaction Collection , No. A list of timestamps for each snapshot interaction and The student feature matrix for each snapshot composition.

[0054] S2. Take the original nodes in S1 as motif-enhanced nodes, retain the original edges, and take each CSCL group as a virtual hyperedge. After each virtual hyperedge connects the specified motif-enhanced nodes, a motif-enhanced graph view is obtained. The motif-enhanced graph view is transformed into a dual hypergraph to obtain a motif-enhanced dual hypergraph view.

[0055] It should be noted that step S2 of the present invention is the construction of a motif-enhanced dual hypergraph view. Most previous anomaly detection works use graph neural networks such as GCN to learn node representations for anomaly detection. However, the groups in CSCL are high-order structures. Treating the groups as a whole and learning the representation of the groups is crucial for accurately detecting group anomalies. Therefore, the present invention proposes a motif-enhanced dual hypergraph view to learn the complex interactions between groups, groups, and students. When constructing a motif-enhanced dual hypergraph view, first, the present invention creates a motif-enhanced graph view by taking the original node as a motif-enhanced node, retaining the original edge, and abstracting the motif instance (CSCL group) as a virtual hyperedge. Since in addition to the students themselves, the interactions between students also contain a lot of available information, and considering that the hypergraph neural network has the ability to encode high-order relationships, the motif-enhanced graph view is further subjected to a dual hypergraph transformation (DHT), that is, the nodes in the motif-enhanced graph view (i.e., the motif-enhanced nodes) are converted into hyperedges, and the edges in the motif-enhanced graph view are converted into supernodes, such as Figure 2 As shown in the upper left corner, a hypergraph neural network can be used to learn the representation of each student and the interactions between students.

[0056] In this embodiment, the specific process of constructing the motif enhancement graph view is as follows: by searching for the CSCL group set in the snapshot of the student interaction in the online classroom , represents the number of CSCL groups, Indicates of snapshots CSCL groups, Represents the index / serial number of the CSCL group. Each CSCL group in the CSCL group set is regarded as a virtual hyperedge, that is, Each virtual hyperedge must connect all original nodes in the corresponding group and the first-order neighbors of all original nodes in the corresponding group (that is, all original nodes that have had direct interactions with the corresponding group).

[0057] S3. Obtain the temporal features of all supernodes in the motif-enhanced dual hypergraph view from the student online classroom interaction data to be tested, take the supernode obtained by the original edge as an ordinary supernode, and take the supernode obtained by the virtual hyperedge as a motif supernode, and stack the temporal feature vectors of the ordinary supernode and the temporal feature vectors of the motif supernode into a matrix form, respectively, to obtain the temporal feature matrix of the ordinary supernode and the temporal feature matrix of the motif supernode, and concatenate the two temporal feature matrices to obtain a complete temporal feature matrix.

[0058] It should be noted that step S3 of the present invention is the encoding and decoding of the student grouping time sequence information. The interaction between students contains fine-grained time information, and the use of a powerful encoder-decoder framework can achieve fine-grained time sequence level anomaly detection. For convenience, in this embodiment, it is taken that each CSCL group contains 3 group members that need to be detected, such as Figure 2 As shown in the figure, we first need to obtain the temporal features of all supernodes in the base-order enhanced dual hypergraph view based on the student online classroom interaction data to be detected. The temporal features of all supernodes obtained from the original edges are , the temporal characteristics of the motif supernode obtained by the virtual hyperedge are , the supernode obtained by the original edge is used as an ordinary supernode, and the supernode obtained by the virtual superedge is used as a base supernode. The time series feature vectors of the ordinary supernode and the time series feature vectors of the base supernode are stacked into a matrix form, and two corresponding time series feature matrices are obtained. Finally, the two time series feature matrices are spliced ​​to obtain a complete time series feature matrix .in, Indicates the snapshot number. is the timestamp corresponding to the ordinary super node, is the timestamp of the motif supernode.

[0059] S4. Input the motif-enhanced dual hypergraph view and the complete temporal feature matrix into the trained anomaly grouping detection model, and output the total anomaly score of the students and the total anomaly score of the CSCL groups or interactions. The total anomaly score of the CSCL groups or interactions includes the CSCL group anomaly score and the anomaly score of the interaction between students.

[0060] It should be noted that step S4 of the present invention constructs an abnormal grouping detection model. In the above-mentioned abnormal grouping detection model, the encoder of the timing module encodes the complete timing feature matrix, and the decoder of the timing module decodes the encoding result of the encoder, and the encoding result of the decoder is used as the timing reconstruction matrix of each snapshot, and the hyperedge features in the motif-enhanced dual hypergraph view are used as the student's own features, and the encoding result of the encoder and the randomly initialized supernode features are weighted and summed as the supernode features of each snapshot, and the student's own features with random values ​​and the supernode features of each snapshot are input into the hypergraph joint neural network of the hypergraph module, and the hypergraph joint neural network is composed of a hyperedge convolutional network and a supernode convolutional network. The hyperedge convolution network and the supernode convolution network have the same number of layers, the output of each layer of the hyperedge convolution network and the output of the corresponding layer of the supernode convolution network are used to calculate the output of the next layer of the hyperedge convolution network, the output of each layer of the supernode convolution network and the output of the corresponding layer of the hyperedge convolution network are used to calculate the output of the next layer of the supernode convolution network, the supernode features of the last layer of the supernode convolution network are used as the encoding of the interactive information by the hypergraph joint neural network, the hyperedge features of the last layer of the hyperedge convolution network are used as the encoding of the student information by the hypergraph joint neural network, and the information to be processed is obtained from the encoding of the student information by the hypergraph joint neural network. The current snapshot information is processed into high-frequency scale and low-frequency scale with the timestamp before the current snapshot information. The high-frequency scale and the low-frequency scale are respectively operated by the multi-head self-attention mechanism. The high-frequency result and the low-frequency result obtained by the multi-head self-attention mechanism are spliced ​​as the student information encoding containing time information. The student information encoding containing time information and the super-node features output by the last layer of the super-node convolutional network are calculated by the self-attention mechanism with Token for each CSCL group to obtain the encoding features of each CSCL group in each snapshot. The encoding features of all CSCL groups in each snapshot are compared with the super-node The super-node features output by the last layer of the point convolutional network are concatenated to obtain the interactive information coding with group information coding. The reconstructed hypergraph association matrix of each snapshot is calculated based on the interactive information coding with group information coding and the student information coding containing time information. The temporal anomaly score is calculated based on the complete temporal feature matrix of each snapshot and the temporal reconstruction matrix. The hybrid anomaly score is calculated based on the association matrix of the motif-enhanced dual hypergraph view and the reconstructed hypergraph association matrix of each snapshot. The total anomaly score of the CSCL group or interaction is calculated by weighted summing the temporal anomaly score and the hybrid anomaly score. The latter in the total anomaly score of the CSCL group or interaction is added. The item is used as the CSCL group anomaly score, and the remaining items are used as the student-to-student interaction anomaly scores. The total anomaly score of the students is calculated based on the association matrix of the motif-augmented dual hypergraph view and the reconstructed hypergraph association matrix of each snapshot.

[0061] It should be noted that step S4 of the present invention constructs an abnormal group detection model. The following describes in detail the structures of various parts of the abnormal group detection model of the present invention and the specific processing flow.

[0062] In order to realize the encoding and decoding of student group timing information, the present invention designs a timing module in the abnormal group detection model. The encoder of the timing module encodes the complete timing feature matrix in units of snapshots, and the decoder of the timing module decodes the encoding result of the encoder to obtain the timing reconstruction matrix of each snapshot.

[0063] In this embodiment, the encoder and the decoder are both implemented using MLP. Specifically, the encoder is first used to encode the complete temporal feature matrix, and then the decoder is used to decode the encoder encoding result:

[0064]

[0065]

[0066] in, For the The encoding result of the complete temporal feature matrix of snapshots; is the activation function; represents the weight of the encoder; Indicates the encoder bias; represents the weight of the decoder; represents the bias of the decoder; is the encoding result of the decoder, i.e. The time series reconstruction matrix of the snapshots.

[0067] After obtaining the time reconstruction matrix of each snapshot, the hypergraph joint neural network based on supernodes and hyperedges encodes the student and interaction messages. Specifically, the hypergraph neural network performs CSCL anomaly grouping detection, focusing on making full use of the information of the online classroom to obtain the anomaly score of the grouping. Student information, interaction information, and time information are all important information that cannot be ignored. The present invention uses the hyperedge features in the motif-enhanced dual hypergraph view as the student's own features. Since the features of interactions between students in large-scale online teaching are difficult to obtain, the features of interactions between students are initially random values. Considering that the encoding result of the encoder is the encoding of the supernode timing information, the encoding result of the encoder is With randomly initialized supernode features The weighted sum is used as the supernode feature of each snapshot ,in, represents weighted sum. The student's own characteristics that will eventually take random values And the supernode characteristics of each snapshot The information is input into the hypergraph joint neural network of the hypergraph module to simultaneously encode student information, interaction information and fine-grained temporal information.

[0068] The hypergraph joint neural network in the present invention is described in detail below. The hypergraph joint neural network includes a supernode convolutional network and a hyperedge convolutional network. Traditional hypergraph neural networks do not consider information from hyperedges when performing convolution operations on supernodes, resulting in limited effectiveness of learning representations. In order to make full use of the information and structure of online classes, the present invention proposes to use students' own characteristics as hyperedge information and aggregate them into the supernode convolutional network. Similarly, from the perspective of hyperedges, the present invention follows the design of the supernode convolutional network and designs a hyperedge convolutional network based on the same concept.

[0069] In this embodiment, the super node convolutional network The output of the layer is specifically expressed as follows:

[0070]

[0071] in, is the activation function; The degree matrix of the supernodes in the motif-augmented dual hypergraph view, The degree matrix of the hyperedges in the dual hypergraph view for motif enhancement, both are used for normalization; An incidence matrix for motif-enhanced dual hypergraph views; is the hyperedge weight matrix, i.e., the importance weight of each student. When the hyperedge weight matrix is ​​difficult to obtain, it is set to the identity matrix; is the first compromise parameter; is the super node convolutional network The output of the layer, i.e., the supernode convolutional network Layer encoding of interaction information between students; Represents the hyperedge convolutional network The output of the layer, i.e., the hyperedge convolutional network Layer encoding of student information; and All are super node convolutional networks. The trainable weight parameters of the layer; Indicates transpose.

[0072] In this embodiment, the hyperedge convolutional network Output of the layer The specific expressions are as follows:

[0073]

[0074] in, is the supernode weight matrix, i.e., the importance weight of each interaction. When the supernode weight matrix is ​​difficult to obtain, it is set to the identity matrix; is the second compromise parameter; and All are hyperedge convolutional networks. Trainable weight parameters of the layer.

[0075] In this embodiment, the supernode feature of the last layer of the supernode convolutional network is finally output. As the encoding of the interactive information by the hypergraph joint neural network, the hyperedge features output by the last layer of the hyperedge convolutional network are As the hypergraph joint neural network encoding of student information. Indicates the number of layers of the hyperedge convolutional network or supernode convolutional network.

[0076] Since the students included in the same online class at different moments are the same, the same student at different moments may participate in different groups and have different interactions. Therefore, when processing the student encoding of each snapshot, it is necessary to consider the historical information of the student at the previous moment. At the same time, the time intervals between different snapshots and the current moment are different, so it is also necessary to consider the time information contained in each snapshot. The present invention proposes to add time series information to the self-attention mechanism to consider the historical information of past snapshots when performing CSCL abnormal grouping detection in the current class, and then designs a time series self-attention mechanism for cross-class student feature learning.

[0077] In the present invention, the current snapshot information to be processed is obtained from the encoding of the student information by the hypergraph joint neural network, and the historical snapshot information with a timestamp before the current snapshot information is used as the coarse-grained time series information. In order to achieve more comprehensive utilization of coarse-grained time series information, the historical snapshot information is processed into two scales, high frequency and low frequency, and multi-head self-attention operations are performed on each scale. Specifically, in this embodiment, The historical snapshot information of a snapshot can be recorded as , Respectively represent First, the historical snapshot information of the The historical snapshot information of each snapshot is time-series weighted attenuated, and the high-frequency scale is obtained after time-series weight attenuation. , after the high-frequency scale is pooled, the low-frequency scale is obtained ,in, Represents element-wise multiplication; is the time series weight matrix, the first The components can be expressed as:

[0078]

[0079] in, Indicates the timestamp of historical snapshot information.

[0080] The high-frequency scale and the low-frequency scale are each passed through a multi-head self-attention mechanism to obtain high-frequency results and low-frequency results respectively. Finally, the high-frequency results and the low-frequency results are concatenated as the first A snapshot contains student information encoding of time information :

[0081]

[0082] in, Represents a splicing operation; and Each represents a multi-head self-attention mechanism.

[0083] In this embodiment, Figure 2 The "HiLo" in refers to the two multi-head self-attention mechanisms for processing high-frequency scales and low-frequency scales. In addition, taking the first snapshot as an example, the historical snapshot information of the first snapshot is the information encoded by itself. .

[0084] In order to obtain a discriminative group instance representation with temporal information, an effective method is needed to aggregate features of students and interactions from group instances. Considering that the self-attention mechanism can adaptively aggregate features and has been widely used to generate representations of the entire sequence, the present invention adds a Token to the head of each sequence. After the self-attention mechanism calculates, the output representation of the Token is fed to the downstream task as the group representation, and a self-attention mechanism with Token is designed. Specifically, in the present invention, the specific calculation method of the encoding features of each CSCL group in each snapshot calculated by the self-attention mechanism with Token is as follows:

[0085] The token that will be used to save the feature aggregation results Encoding student information including time information Middle The hyperedge features of the CSCL groups and the supernode features output by the last layer of the supernode convolutional network Middle The features of each supernode of the CSCL group are concatenated and calculated through the self-attention mechanism to obtain the output feature. The 0th element of the output feature is used as the first element in each snapshot. Coding characteristics of the CSCL groups.

[0086] Furthermore, in this embodiment, the specific calculation method for obtaining the coding features of each CSCL group in each snapshot is as follows:

[0087] ;

[0088] in, In each snapshot, Coding characteristics of the CSCL groups; Represents the calculation of the self-attention mechanism; Represents the token used to save feature aggregation results; Represents the 0th element of the output feature; Respectively represent the student information code containing time information Middle CSCL group Hyperedge features; Respectively represent the super node features output by the last layer of the super node convolutional network Middle CSCL group Supernode features; Represents the number of members in each CSCL group.

[0089] Finally, the encoding features of all CSCL groups in each snapshot are combined with the supernode features output by the last layer of the supernode convolutional network. Concatenate to get interactive information coding with group information coding Among them, The mutual information coding with group information coding for snapshots can be expressed as In this embodiment, taking the first snapshot as an example, the corresponding student information encoding containing time information can be expressed as , where the feature of a hyperedge numbered a can be expressed as .

[0090] After obtaining the fine-grained timing module The time reconstruction matrix of snapshots And use the hypergraph to combine neural network and self-attention mechanism to obtain the The snapshot captures the graph structure and temporal information of the student information encoding and mutual information coding with group information coding . Then it can be encoded by mutual information with group information coding and student information code containing time information Compute the reconstructed hypergraph association matrix for each snapshot :

[0091]

[0092] in, Represents the sigmoid function.

[0093] The timing anomaly score is calculated based on the complete timing feature matrix of each snapshot and the timing reconstruction matrix of each snapshot. Specifically, the timing reconstruction matrix of each CSCL group or interaction in each snapshot is subtracted from the corresponding complete timing feature matrix to obtain the first subtraction result, and the L2 norm of the first subtraction result is used as the timing anomaly score of each CSCL group or interaction:

[0094]

[0095] in, Indicates the supernode number corresponding to the CSCL group or interaction; Indicates The temporal anomaly scores of the CSCL groups or interactions; Indicates Snapshot A time-reconstruction matrix for a CSCL group or interaction; Indicates Snapshot A complete temporal feature matrix for each CSCL group or interaction; represents the L2 norm.

[0096] The hybrid anomaly score is calculated based on the association matrix of the motif-enhanced dual hypergraph view and the reconstructed hypergraph association matrix of each snapshot. Specifically, the reconstructed hypergraph association matrix of each CSCL group or interaction in each snapshot is subtracted from the association matrix of the corresponding motif-enhanced dual hypergraph view to obtain the second subtraction result, and the L2 norm of the second subtraction result is used as the hybrid anomaly score of each CSCL group or interaction:

[0097]

[0098] in, Indicates Snapshot The association matrix of the motif-enhanced dual hypergraph view of CSCL groups or interactions; Indicates Snapshot The reconstructed hypergraph association matrix of each CSCL group or interaction.

[0099] The total anomaly score of the CSCL group or interaction is calculated by weighted summing the time series anomaly score and the mixed anomaly score:

[0100]

[0101] in, represents a hyperparameter; represents the total anomaly score of the CSCL group or interaction, where Items represent CSCL group abnormality scores , and the remaining items represent the abnormal scores of interactions between students During the inference of the anomaly group detection model, the CSCL group anomaly score is output as the most important item.

[0102] The total anomaly score of the student is calculated based on the association matrix of the motif-enhanced dual hypergraph view and the reconstructed hypergraph association matrix of each snapshot. Specifically, the reconstructed hypergraph association matrix of each student in each snapshot is subtracted from the association matrix of the corresponding motif-enhanced dual hypergraph view to obtain the third subtraction result, and the L2 norm of the third subtraction result is used as the total anomaly score of each student:

[0103]

[0104] in, Indicates the student's serial number; Indicates Total abnormal score of students; Indicates Snapshot The reconstructed hypergraph association matrix of each student; Indicates Snapshot The incidence matrix of the student's motif-augmented dual hypergraph view.

[0105] It should be noted that in the training process of the anomaly grouping detection model, the time series anomaly score is used as the reconstruction loss of the complete time series feature matrix. , using the mixed anomaly score as the reconstruction loss of the association matrix of the motif-enhanced dual hypergraph view The total loss of the anomaly group detection model is taken as the weighted sum of the reconstruction loss of the complete temporal feature matrix and the reconstruction loss of the association matrix of the motif-enhanced dual hypergraph view. , based on minimizing the total loss, the parameters of the anomaly group detection model are updated until the total loss converges to obtain a trained anomaly group detection model.

[0106] In this embodiment, the reconstruction loss of the complete time series feature matrix is Specifically:

[0107]

[0108] in, Indicates The time-series reconstruction matrix of snapshots; Indicates The complete time series feature matrix for each snapshot.

[0109] In this embodiment, the motif enhances the reconstruction loss of the association matrix of the dual hypergraph view Specifically:

[0110]

[0111] in, Indicates The snapshot motifs enhance the incidence matrix of the dual hypergraph view; Indicates The reconstructed hypergraph incidence matrix of the snapshots.

[0112] In this embodiment, the total loss Specifically:

[0113]

[0114] It should be noted that the total loss for each training The error back propagation algorithm and stochastic gradient descent are used to train the anomaly group detection model.

[0115] It should also be noted that in the anomaly group detection of computer-supported collaborative learning based on a hypergraph network, the above-mentioned anomaly group detection model is first trained, and then a large-scale video of students learning online is found as input data. Whenever a CSCL discussion session is entered, the interaction data between students is collected, and it is converted into a motif-enhanced dual hypergraph view according to the process of the present invention to obtain a complete temporal feature matrix. Then the motif-enhanced dual hypergraph view and the complete temporal feature matrix are input into the trained anomaly group detection model to obtain the fine-grained temporal information and mixed information encoding of the discussion in the online class, that is, the temporal reconstruction matrix of all snapshots output by the timing module is obtained. As well as the interactive information coding with group information coding and student information coding containing time information output by the hypergraph module Next, based on the encoding information obtained by the above anomaly grouping detection model, the total anomaly score of students in the online class, the anomaly score of the CSCL group, and the anomaly score of student interactions are calculated based on the reconstruction error, based on the assumption that the anomaly content is more difficult to reconstruct. ,Obtaining accurate anomaly scores for each CSCL group is the main goal of the ,presentation, and finally outputting the anomaly scores of the CSCL groups in the online ,classroom and supplemented by multi-faceted explanations.

[0116] The present invention constructs the student online classroom interaction data into a discrete dynamic graph and obtains the CSCL group abnormality score through the abnormal grouping detection model. At the same time, the structural abnormality degree, the temporal abnormality degree and the attribute abnormality degree of each student interaction obtained in the process are supplemented to provide a certain degree of explainability, so as to assist teachers in timely and accurately discovering the grouping of problems and accurately locating the crux of the problem in the large-scale online classroom CSCL process. Therefore, the present invention fully considers the rich information contained in the online classroom and uses dynamic graph modeling to perform CSCL grouping abnormality detection in education, which is more convincing, and can also promote the development of large-scale online learning and improve the effect of CSCL.

[0117] The present invention will now use a specific example to demonstrate the application effect of the hypergraph-based computer-supported collaborative learning anomaly grouping detection method described in S1 to S4 in the above embodiments on a specific data set, so as to facilitate understanding of the essence of the present invention.

[0118] Example

[0119] The specific implementation process of the abnormal grouping detection method for computer-supported collaborative learning based on hypergraph adopted in this embodiment is as described above and will not be repeated here. In order to demonstrate the technical effect of the method proposed in the present invention, it is verified in an actual teaching scenario. The data set selected by the present invention is an educational data set, which is collected from a real large-scale learning scenario supported by a computer. A reasonable number of group discussions will be carried out during the course, and the teacher will arrange reasonable grouping in advance for each group discussion in the lesson preparation link. In addition, in order to verify the effectiveness of the method proposed in the present invention, a small number of groups that do not meet the grouping principles such as homogeneity between groups and heterogeneity within groups or randomly disrupted groups will be arranged in the classroom. The personal basic information of students in the classroom is collected before the course starts as the initial features of each student, and the interaction between students is required to be implemented by computer for real-time collection. In this data set, each node represents a student in the large-scale classroom, and each edge represents the communication between two students in the classroom.

[0120] The most commonly used AUC in the field of anomaly detection is selected as the evaluation index. It is the area under the ROC curve, ranging from 0 to 1. The larger the AUC value, the better the model effect. The comparison results between the method proposed in this invention and the benchmark method for dynamic graph anomaly detection are shown in Table 1. For the sake of fairness, subsequent experiments are carried out under the above-mentioned education dataset with an anomaly rate of 10%.

[0121] Table 1. Results of abnormal group detection using different methods on the education dataset

[0122]

[0123] The experimental results in Table 1 show that the method proposed in the present invention has a significant improvement over the baseline method. One main reason is that the method proposed in the present invention considers each grouping instance from a holistic perspective. The above-mentioned comparison methods are: Netwalk, a deep embedding dynamic graph anomaly detection model, which comes from the existing technical literature: W. Yu, W.Cheng, CC Aggarwal, K. Zhang, H. Chen, and W. Wang, “NetWalk: A flexible deep embedding approach for anomaly detection in dynamic networks,” in Proc.24th ACM SIGKDD Int. Conf. Knowl. Discovery Data Mining, Jul. 2018, pp. 2672–2681; AddGraph, a dynamic graph anomaly detection model based on attention and timing, which comes from the existing technical literature: L.Zheng, Z. Li, J. Li, Z. Li, and J. Gao, “AddGraph: Anomaly detection indynamic graph using attention-based temporal GCN,” in Proc. IJCAI, Aug. 2019, pp. 4419–4425; TADDY, a dynamic graph anomaly detection model based on Transformer, which comes from the existing technical literature: Y. Liu et al., “Anomaly Detection in dynamic graphs via transformer,” IEEETrans. Knowl. Data Eng., early access, Nov. 2, 2021, doi: 10.1109 / TKDE.2021.3124061.

[0124] In addition, this embodiment also conducts an ablation experiment to better examine the contribution of the innovation proposed by the present invention, and the experimental results are shown in Table 2. In the ablation experiment, the specific details of each module are as follows: Backbone: Use the graph convolutional network (GCN) to perform node-level representation learning on the original dynamic graph corresponding to the data set, and use average aggregation as the aggregation method for group representation; W / Hyper: On the basis of Backbone, the original dynamic graph structure is formed into a base-enhanced dual hypergraph view by adding virtual hyperedges and performing dual hypergraph transformation, and a hypergraph joint network is used for hypergraph representation learning, and finally a self-attention mechanism with Token is used to aggregate group features; W / Time: On the basis of W / Hyper, a timing module and a timing self-attention mechanism are added, corresponding to the complete model of the present invention.

[0125] Table 2. Ablation test results of the innovative module of the present invention on the education dataset

[0126]

[0127] The results of the ablation experiment show that the two innovative modules proposed in the present invention have significant effects. The main reasons are: the hypergraph module enables the abnormal group detection model to encode the high-order relationship of the group and adaptively aggregate the group representation; the timing module can simultaneously consider the multi-granularity time information of the dynamic graph, such as the time information contained in the classroom. In summary, the present invention can provide an efficient and accurate abnormal group detection solution for unsupervised computer-supported collaborative learning scenarios.

[0128] It should also be noted that the hypergraph-based computer-supported collaborative learning anomaly grouping detection method in the above embodiment can essentially be executed by a computer program or module. Therefore, similarly, based on the same inventive concept, another preferred embodiment of the present invention also provides a hypergraph-based computer-supported collaborative learning anomaly grouping detection system corresponding to the hypergraph-based computer-supported collaborative learning anomaly grouping detection method provided in the above embodiment, such as Figure 3 As shown, it includes:

[0129] The data acquisition and processing module is used to obtain the online classroom interaction data of students to be tested, and represents students as original nodes and the interactions between students as original edges. When three or more original nodes are fully connected, it is considered that there is dense interaction between students. Students with dense interaction form a CSCL group. All interactions are divided into several snapshots according to the time sequence, and all snapshots form a discrete original dynamic graph structure; each snapshot consists of its own student set, interaction set, interaction timestamp list and student feature matrix;

[0130] A motif-enhanced dual hypergraph view acquisition module is used to use the original nodes in the data acquisition and processing module as motif-enhanced nodes, retain the original edges, and use each CSCL group as a virtual hyperedge. After each virtual hyperedge connects the specified motif-enhanced node, a motif-enhanced graph view is obtained. The motif-enhanced graph view is subjected to a dual hypergraph transformation to obtain a motif-enhanced dual hypergraph view.

[0131] A complete temporal feature matrix acquisition module is used to obtain the temporal features of all supernodes in the motif-enhanced dual hypergraph view from the student online classroom interaction data to be detected, and the supernodes obtained from the original edges are used as ordinary supernodes, and the supernodes obtained from the virtual hyperedges are used as motif supernodes. The temporal feature vectors of the ordinary supernodes and the temporal feature vectors of the motif supernodes are stacked into matrix form respectively, and the temporal feature matrices of the ordinary supernodes and the temporal feature matrices of the motif supernodes are obtained accordingly. The two temporal feature matrices are spliced ​​to obtain a complete temporal feature matrix;

[0132] The result acquisition module is used to input the motif-enhanced dual hypergraph view and the complete temporal feature matrix into the trained anomaly grouping detection model, and output the total anomaly score of the students and the total anomaly score of the CSCL group or interaction. The total anomaly score of the CSCL group or interaction includes the CSCL group anomaly score and the anomaly score of the interaction between students.

[0133] The above-described embodiment is only a preferred solution of the present invention, but it is not intended to limit the present invention. A person skilled in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, any technical solution obtained by equivalent replacement or equivalent transformation falls within the protection scope of the present invention.

Claims

1. A method for abnormal grouping detection based on hypergraph computer supported collaborative learning, characterized in that: The following steps are involved: S1. Obtain the online classroom interaction data of students to be tested, represent students as original nodes, and use the interactions between students as original edges. When three or more original nodes are fully connected, it is considered that there is dense interaction between students. Students with dense interactions form a CSCL group. All interactions are divided into several snapshots according to the time sequence, and all snapshots form a discrete original dynamic graph structure; each snapshot consists of its own student set, interaction set, interaction timestamp list and student feature matrix; S2. Take the original nodes in S1 as motif-enhanced nodes, keep the original edges, and take each CSCL group as a virtual hyperedge. After each virtual hyperedge connects the specified motif-enhanced nodes, a motif-enhanced graph view is obtained. The motif-enhanced graph view is transformed into a dual hypergraph to obtain a motif-enhanced dual hypergraph view. S3. Obtain the temporal features of all supernodes in the motif-enhanced dual hypergraph view from the student online classroom interaction data to be detected, take the supernode obtained by the original edge as the ordinary supernode, and take the supernode obtained by the virtual hyperedge as the motif supernode, respectively stack the temporal feature vectors of the ordinary supernode and the temporal feature vectors of the motif supernode into a matrix form, and obtain the temporal feature matrix of the ordinary supernode and the temporal feature matrix of the motif supernode, and concatenate the two temporal feature matrices to obtain a complete temporal feature matrix; S4. Input the motif-enhanced dual hypergraph view and the complete temporal feature matrix into the trained anomaly grouping detection model, and output the total anomaly score of the students and the total anomaly score of the CSCL groups or interactions. The total anomaly score of the CSCL groups or interactions includes the CSCL group anomaly score and the anomaly score of the interaction between students.

2. A method for detecting abnormal grouping in computer-supported collaborative learning based on a hypergraph as claimed in claim 1, characterized in that: In the abnormal group detection model, the encoder of the timing module encodes the complete timing feature matrix, and the decoder of the timing module decodes the encoding result of the encoder, and the encoding result of the decoder is used as the timing reconstruction matrix of each snapshot, and the hyperedge features in the base-enhanced dual hypergraph view are used as the student's own features, and the encoding result of the encoder and the randomly initialized supernode features are weighted and summed as the supernode features of each snapshot, and the student's own features with random values ​​and the supernode features of each snapshot are input into the hypergraph joint neural network of the hypergraph module, and the hypergraph joint neural network is composed of a hyperedge convolutional network and a supernode convolutional network. The hyperedge convolution network and the supernode convolution network have the same number of layers, the output of each layer of the hyperedge convolution network and the output of the corresponding layer of the supernode convolution network are used to calculate the output of the next layer of the hyperedge convolution network, the output of each layer of the supernode convolution network and the output of the corresponding layer of the hyperedge convolution network are used to calculate the output of the next layer of the supernode convolution network, the supernode features of the last layer of the supernode convolution network are used as the encoding of the interactive information by the hypergraph joint neural network, the hyperedge features of the last layer of the hyperedge convolution network are used as the encoding of the student information by the hypergraph joint neural network, and the information to be processed is obtained from the encoding of the student information by the hypergraph joint neural network. The current snapshot information is processed into high-frequency scale and low-frequency scale with the timestamp before the current snapshot information. The high-frequency scale and the low-frequency scale are respectively operated by the multi-head self-attention mechanism. The high-frequency result and the low-frequency result obtained by the multi-head self-attention mechanism are spliced ​​as the student information encoding containing time information. The student information encoding containing time information and the super-node features output by the last layer of the super-node convolutional network are calculated by the self-attention mechanism with Token for each CSCL group to obtain the encoding features of each CSCL group in each snapshot. The encoding features of all CSCL groups in each snapshot are compared with the super-node The super-node features output by the last layer of the point convolutional network are concatenated to obtain the interactive information coding with group information coding. The reconstructed hypergraph association matrix of each snapshot is calculated based on the interactive information coding with group information coding and the student information coding containing time information. The temporal anomaly score is calculated based on the complete temporal feature matrix of each snapshot and the temporal reconstruction matrix. The hybrid anomaly score is calculated based on the association matrix of the motif-enhanced dual hypergraph view and the reconstructed hypergraph association matrix of each snapshot. The total anomaly score of the CSCL group or interaction is calculated by weighted summing the temporal anomaly score and the hybrid anomaly score. The latter in the total anomaly score of the CSCL group or interaction is added. The item is used as the CSCL group anomaly score, and the remaining items are used as the student-to-student interaction anomaly scores. The total anomaly score of the student is calculated based on the association matrix of the motif-enhanced dual hypergraph view and the reconstructed hypergraph association matrix of each snapshot; where, is the number of CSCL groups.

3. A method for detecting abnormal grouping in computer-supported collaborative learning based on a hypergraph as claimed in claim 2, characterized in that: Supernode Convolutional Network Output of the layer The specific expressions are as follows: ； in, is the activation function; Degree matrices of supernodes in the dual hypergraph view are enhanced for motifs; An incidence matrix for motif-enhanced dual hypergraph views; is the hyperedge weight matrix; Degree matrix of hyperedges in dual hypergraph view for motif enhancement; is the first compromise parameter; is the super node convolutional network The output of the layer; Represents the hyperedge convolutional network The output of the layer; and All are super node convolutional networks. The trainable weight parameters of the layer; Indicates transpose.

4. A method for detecting abnormal grouping in computer-supported collaborative learning based on a hypergraph as claimed in claim 3, characterized in that: Hyperedge Convolutional Network Output of the layer The specific expressions are as follows: ； in, is the supernode weight matrix; is the second compromise parameter; and All are hyperedge convolutional networks. Trainable weight parameters of the layer.

5. The method for abnormal grouping detection based on hypergraph computer supported collaborative learning as claimed in claim 2, characterized in that: The high-frequency scale and the low-frequency scale are obtained as follows: historical snapshot information with a timestamp before the current snapshot information is subjected to time series weight decay to obtain a high-frequency scale after time series weight decay, and the high-frequency scale is subjected to a pooling operation to obtain a low-frequency scale.

6. A method for abnormal grouping detection based on hypergraph computer supported collaborative learning as claimed in claim 2, characterized in that: Get the first The specific calculation method of the encoding features of the CSCL group is as follows: the token used to save the feature aggregation result Encoding student information including time information Middle The hyperedge features of the CSCL groups and the supernode features output by the last layer of the supernode convolutional network Middle The features of each supernode of the CSCL group are concatenated and calculated through the self-attention mechanism to obtain the output feature. The 0th element of the output feature is used as the first element in each snapshot. Coding characteristics of the CSCL groups.

7. A method for detecting abnormal grouping in computer-supported collaborative learning based on a hypergraph as claimed in claim 2, characterized in that: In step S2, the timing reconstruction matrix of each CSCL group or interaction in each snapshot is subtracted from the corresponding complete timing feature matrix to obtain a first subtraction result, and the L2 norm of the first subtraction result is used as the timing anomaly score of each CSCL group or interaction; the reconstructed hypergraph association matrix of each CSCL group or interaction in each snapshot is subtracted from the association matrix of the corresponding motif-enhanced dual hypergraph view to obtain a second subtraction result, and the L2 norm of the second subtraction result is used as the mixed anomaly score of each CSCL group or interaction; the reconstructed hypergraph association matrix of each student in each snapshot is subtracted from the association matrix of the corresponding motif-enhanced dual hypergraph view to obtain a third subtraction result, and the L2 norm of the third subtraction result is used as the total anomaly score of each student.

8. The method for detecting abnormal grouping in computer-supported collaborative learning based on a hypergraph as claimed in claim 2, characterized in that: During the training process of the anomaly grouping detection model, the temporal anomaly score is used as the reconstruction loss of the complete temporal feature matrix, the mixed anomaly score is used as the reconstruction loss of the association matrix of the motif-enhanced dual hypergraph view, and the weighted sum of the reconstruction loss of the complete temporal feature matrix and the reconstruction loss of the association matrix of the motif-enhanced dual hypergraph view is used as the total loss of the anomaly grouping detection model. The parameters of the anomaly grouping detection model are updated based on minimizing the total loss until the total loss converges to obtain a trained anomaly grouping detection model.

9. The method for abnormal grouping detection based on hypergraph computer supported collaborative learning according to claim 1, characterized in that: In step S2, the specific process of constructing the motif-enhanced graph view is as follows: a set of CSCL groups is obtained by searching in all snapshots, and each CSCL group in the set of CSCL groups is used as a virtual hyperedge. Each virtual hyperedge connects all the original nodes in the corresponding group and the first-order neighbors of all the original nodes in the corresponding group.

10. A hypergraph-based computer-supported collaborative learning anomaly grouping detection system, characterized in that: include: The data acquisition and processing module is used to obtain the online classroom interaction data of students to be tested, and represents students as original nodes and the interactions between students as original edges. When three or more original nodes are fully connected, it is considered that there is dense interaction between students. Students with dense interaction form a CSCL group. All interactions are divided into several snapshots according to the time sequence, and all snapshots form a discrete original dynamic graph structure; each snapshot consists of its own student set, interaction set, interaction timestamp list and student feature matrix; A motif-enhanced dual hypergraph view acquisition module is used to use the original nodes in the data acquisition and processing module as motif-enhanced nodes, retain the original edges, and use each CSCL group as a virtual hyperedge. After each virtual hyperedge connects the specified motif-enhanced node, a motif-enhanced graph view is obtained. The motif-enhanced graph view is subjected to a dual hypergraph transformation to obtain a motif-enhanced dual hypergraph view. A complete temporal feature matrix acquisition module is used to obtain the temporal features of all supernodes in the motif-enhanced dual hypergraph view from the student online classroom interaction data to be detected, and the supernodes obtained from the original edges are used as ordinary supernodes, and the supernodes obtained from the virtual hyperedges are used as motif supernodes. The temporal feature vectors of the ordinary supernodes and the temporal feature vectors of the motif supernodes are stacked into matrix form respectively, and the temporal feature matrices of the ordinary supernodes and the temporal feature matrices of the motif supernodes are obtained accordingly. The two temporal feature matrices are spliced ​​to obtain a complete temporal feature matrix; The result acquisition module is used to input the motif-enhanced dual hypergraph view and the complete temporal feature matrix into the trained anomaly grouping detection model, and output the total anomaly score of the students and the total anomaly score of the CSCL group or interaction. The total anomaly score of the CSCL group or interaction includes the CSCL group anomaly score and the anomaly score of the interaction between students.

Images