Anomaly Learner Detection Method Based on Graph Aggregation and Recovery Supported by Granular Computing

Abstract

The present invention discloses an abnormal learner detection method based on graph aggregation and restoration supported by granular computing, which can be applied to the field of graph aggregation technology. The method of the present invention includes the following steps: obtaining learner data of an online learning platform; performing data granulation on the learner data to obtain a learner relationship graph; inputting the learner relationship graph into a learner anomaly detection model to predict a learning anomaly detection result of the learner; wherein, the learner anomaly detection model is constructed through the following steps: constructing an aggregation layer by using a graph algorithm based on modularity; constructing a restoration layer by using a skip connection algorithm; constructing an output layer based on the aggregation layer and the restoration layer. The present invention can combine the influence of learner groups at different granularities on anomaly detection, and effectively improves the accuracy of anomaly detection.

Description

Technical Field

[0001] The present invention relates to the technical field of graph aggregation, and in particular to an abnormal learner detection method based on graph aggregation and recovery under the support of granular computing. Background Art

[0002] In related technologies, with the popularization of the Internet, the online learning industry has been booming, and the user group learning through the network has also grown larger. However, compared with the classroom learning mode, due to the lack of direct face-to-face interaction between teachers and students, the learning efficiency and quality of online learning highly depend on the detection of abnormal learning states. Currently, online learning anomaly detection methods mainly consider the static characteristics of learners and a single granularity level, and such modeling methods are difficult to consider the influence of learner groups at different granularities on anomaly detection. Summary of the Invention

[0003] The present invention aims to at least solve one of the technical problems existing in the prior art. For this purpose, the present invention proposes an abnormal learner detection method based on graph aggregation and recovery under the support of granular computing, which can effectively improve the accuracy of anomaly detection.

[0004] An embodiment of the present invention provides an abnormal learner detection method based on graph aggregation and recovery under the support of granular computing, including the following steps:

[0005] Obtain learner data of an online learning platform;

[0006] Perform data granulation on the learner data to obtain a learner relationship graph;

[0007] Input the learner relationship graph into a learner anomaly detection model to predict a learning anomaly detection result of the learner;

[0008] Wherein, the learner anomaly detection model is constructed through the following steps:

[0009] Construct an aggregation layer by using a graph algorithm based on modularity;

[0010] Construct a recovery layer by using a skip connection algorithm;

[0011] Construct an output layer based on the aggregation layer and the recovery layer.

[0012] In some embodiments, the performing data granulation on the learner data to obtain a learner relationship graph includes:

[0013] Extract features related to learner anomalies from the learner data and calculate the weights of the features related to learner anomalies;

[0014] Generate a learner relationship graph according to the features related to learner anomalies and the weights.

[0015] In some embodiments, extracting features related to learner anomalies from the learner data and calculating the weights of the features related to learner anomalies includes:

[0016] Extracting features related to learner anomalies from the learner data to obtain a learner feature set, a feature set of features to be determined for importance, and a learner anomaly situation set;

[0017] Substituting the learner feature set, the feature set of features to be determined for importance, and the learner anomaly situation set into rough set theory for calculation to obtain a first positive domain of the learner anomaly situation set under the learner feature set and a second positive domain of the learner anomaly situation set under the feature set of features to be determined for importance;

[0018] Determining the importance definition of the feature set of features to be determined for importance with respect to the learner anomaly situation set according to the first positive domain and the second positive domain;

[0019] Screening the feature set of features to be determined for importance according to the importance definition;

[0020] Calculating the weight of each feature in the screened feature set of features to be determined for importance.

[0021] In some embodiments, generating a learner relationship graph according to the features related to learner anomalies and the weights includes:

[0022] Granulating the features related to learner anomalies according to the types of the features related to learner anomalies;

[0023] Calculating the weight values of the edges of each node in the learner relationship graph according to the granulation result and the weight of each feature;

[0024] Generating a learner relationship graph according to the features related to learner anomalies and the weight values of the edges of each node.

[0025] In some embodiments, after inputting the learner relationship graph into a learner anomaly detection model, the learner anomaly detection model performs the following steps:

[0026] Partitioning the feature nodes in the learner relationship graph through the aggregation layer;

[0027] Adjusting the partitioning process of the aggregation layer through the restoration layer;

[0028] Predicting the label probabilities of the unlabeled nodes in the learner relationship graph through the classifier of the output layer.

[0029] In some embodiments, partitioning the feature nodes in the learner relationship graph through the aggregation layer includes:

[0030] Predict the graph topology and graph representation according to the learner relationship graph, and use graph wavelet convolution to extract hidden feature nodes;

[0031] Define all feature nodes in the learner relationship graph as a community respectively, and initialize the community as an unmarked state. The feature nodes include hidden feature nodes and unhidden feature nodes;

[0032] Calculate the modularity of the preset community;

[0033] Adjust the subordination relationship between the unmarked community and the preset community according to the modularity, and mark the nodes corresponding to the community with the determined subordination relationship;

[0034] Obtain the unmarked nodes, and calculate the normalized weights of the edges of all first-order neighbor nodes of the unmarked nodes;

[0035] Put the unmarked nodes into the community with the highest normalized weight, and mark the unmarked nodes;

[0036] Construct an aggregation graph according to the marked nodes.

[0037] In some embodiments, the adjusting the subordination relationship between the unmarked community and the preset community according to the modularity, and marking the nodes corresponding to the community with the determined subordination relationship includes:

[0038] Determine a preset difference in the modularity for moving the unmarked community to the preset community;

[0039] Determine that the preset difference is greater than zero and the number of nodes in the neighbor community of the nodes corresponding to the preset community is less than a preset number, and move the unmarked community to the preset community;

[0040] Determine that the modularity is in a stable state, and stop the moving process of the unmarked community.

[0041] In some embodiments, the adjusting the partitioning process of the aggregation layer by the restoration layer includes:

[0042] Use the feature transformation of the graph wavelet convolution network and graph convolution to generate the hidden feature nodes in the input graph of the restoration layer;

[0043] According to the hidden feature nodes in the input graph of the restoration layer, use the matching matrix to restore the graph representation of the input graph.

[0044] In some embodiments, the adjusting the partitioning process of the aggregation layer by the restoration layer further includes:

[0045] The hidden feature nodes extracted by using graph wavelet convolution are combined with the hidden feature nodes in the input graph of the recovery layer by using skip connections.

[0046] In some embodiments, the learner anomaly detection model is a pre-trained model, and the training process includes:

[0047] The learner anomaly detection model is globally trained by using a deep learning optimization algorithm.

[0048] An anomaly learner detection method based on graph aggregation and recovery supported by granular computing provided by an embodiment of the present invention has the following beneficial effects:

[0049] In this embodiment, by first granulating learner data to obtain a learner relationship graph, and then inputting the learner relationship graph into a learner anomaly detection model including an aggregation layer, a recovery layer, and an output layer, a learning anomaly detection result of the learner is predicted, so as to combine the influence of learner groups at different granularities on anomaly detection, and the accuracy of anomaly detection is effectively improved.

[0050] Additional aspects and advantages of the present invention will be given in part in the following description, will become apparent in part from the following description, or will be understood through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The present invention will be further described below in conjunction with the drawings and embodiments, where:

[0052] Figure 1 is a flowchart of an anomaly learner detection method based on graph aggregation and recovery supported by granular computing according to an embodiment of the present invention;

[0053] Figure 2 is a flow framework diagram for generating online learner feature particles and graph structures according to an embodiment of the present invention;

[0054] Figure 3 is a schematic framework diagram of a learner anomaly detection model according to an embodiment of the present invention;

[0055] Figure 4 is a data processing flowchart of a learner anomaly detection model according to an embodiment of the present invention;

[0056] Figure 5 is a schematic diagram of the aggregation process of the aggregation layer according to an embodiment of the present invention;

[0057] Figure 6 is a schematic diagram of the recovery process of the recovery layer according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] Embodiments of the present invention will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary only for explaining the present invention and should not be construed as limiting the present invention.

[0059] In the description of the present invention, the meaning of "several" is more than one, the meaning of "multiple" is more than two, and understandings such as "greater than", "less than", "exceeding", etc. do not include the recited number, and understandings such as "above", "below", "within", etc. include the recited number. If there is a description of "first", "second", etc., it is only for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance or implicitly indicating the quantity of the indicated technical features or implicitly indicating the sequence relationship of the indicated technical features.

[0060] In the description of the present invention, unless otherwise clearly defined, words such as "set" should be understood in a broad sense, and those skilled in the art can reasonably determine the specific meaning of the above words in the present invention in combination with the specific content of the technical solution.

[0061] In the description of the present invention, the description with reference to terms such as "one embodiment", "some embodiments", "schematic embodiments", "examples", "specific examples", or "some examples" means that the specific features or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic descriptions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0062] Before elaborating on the specific embodiments, the Chinese meanings represented by the sub-letters or sub-letter combinations involved in the embodiments are explained, and the specific explanations are shown in Table 1:

[0063] Table 1

[0064]

[0065]

[0066] Refer to Figure 1 , an embodiment of the present invention provides an abnormal learner detection method based on graph aggregation and restoration under the support of granular computing. This embodiment can be applied to a server or a background processor of an online learning platform.

[0067] During the application process, this embodiment includes the following steps:

[0068] S11. Obtain learner data of the online learning platform. Among them, the learner data includes the learner's own learning data and the learning data of other learners adjacent to the learner.

[0069] S12. Granulate the learner data to obtain a learner relationship graph.

[0070] In the embodiments of the present application, when using a graph neural network model to detect abnormal learners, the online learners need to be granulated for effective detection. Therefore, in this embodiment, learners are granulated, and the granulated nodes are transformed into a relationship graph. Specifically, in this embodiment, features related to learner anomalies are first extracted from the learner data, and the weights of the features related to learner anomalies are calculated. Then, a learner relationship graph is generated based on the features related to learner anomalies and the weights. For example, as Figure 2 shown, the learners include a, b, c, d, e, and f, and the learner data includes data such as the gender, course completion time, and grades of the learners; after feature selection and extraction of these learner data, the selected features are obtained, and then the weights of these features are calculated. After determining the relationships of these learners based on the weights, an undirected graph is generated as the learner relationship graph.

[0071] Among them, the extraction of features and the calculation of weights can be performed in the following manner:

[0072] Extract features related to learner anomalies from the learner data to obtain a learner feature set, a set of features to be determined for importance, and a set of learner anomaly situations. Then, substitute the extracted feature set into the rough set theory for calculation to obtain the first positive domain of the set of learner anomaly situations under the learner feature set and the second positive domain of the set of learner anomaly situations under the set of features to be determined for importance. Then, based on the first positive domain and the second positive domain, determine the importance definition of the set of features to be determined for importance with respect to the set of learner anomaly situations, and screen the set of features to be determined for importance according to the importance definition. Then, calculate the weight of each feature in the screened set of features to be determined for importance.

[0073] Specifically, in Pawlak rough set theory, for two learners x(i) and x(j), the following can be defined: for any Design its indiscernibility relation where G Rk is the interaction function corresponding to the indiscernibility relation of the Rk class, and ε is the threshold. This enables the present invention to generate a partition U, denoted as U / R AT , = {[x] AT ,|x ∈ U}. [x] AT , is the equivalence class of x with respect to R AT ,. For any X ∈ U, the upper and lower approximations of X are expressed as: Then the positive region, negative region, and boundary region of x are respectively POS(X) = R (x), NEG(X) = U -R (X) and Based on this in this embodiment, bringing in the learner's feature set AT, the feature set AT' whose importance is to be determined, and the learner's abnormal situation set D, the positive domain POS of the abnormal situation set under the learner's feature set can be obtained AT (D) is used as the first positive domain, and the positive domain POS of the abnormal situation set under the feature set whose importance has been determined AT-AT , (D) is used as the second positive domain. And the importance of the feature set AT' to be determined with respect to the learner's abnormal situation set D is defined as shown in the following formula (1):

[0074]

[0075] Among them, the symbol "||" represents the absolute value.

[0076] Then, according to the selected feature set Acc = {AT'|σ D AT’ > ζ}, the attribute AT' is selected, where ζ is a threshold, thereby reducing the features from the original feature set whose importance is to be determined. And the weight of each selected feature set AT' is calculated by formula (2):

[0077]

[0078] Among them, Acc z represents the z-th feature of the selected feature set Acc, and n Acc represents the size of Acc.

[0079] And generating the relationship graph of the learner according to the features and weights can be achieved in the following way:

[0080] Granulate the features related to the learner's abnormality according to the type of the features related to the learner's abnormality, then calculate the weight of the edge of each node in the learner relationship graph according to the granulation result and the weight of each feature, and then generate the learner relationship graph according to the features related to the learner's abnormality and the weight of the edge of each node.

[0081] In this embodiment, for different data types, the granular computing theory is used to granulate the data with different binary relations under the feature subset to obtain the corresponding granularity information and granularity structure. Among them, to obtain the granularity information, this embodiment granulates symbolic data with an equivalence relation, numerical data with an adjacent relation, and incomplete data with a tolerance relation.

[0082] Specifically, for symbolic data, taking AT as the symbolic feature subset of the l-th graph convolutional layer, the indiscernibility relation is as shown in formula (3):

[0083]

[0084] The a-th dimensional feature is denoted as a, which represents the value of the a-th dimension of the i-th learner in the l-th convolutional layer.

[0085] For numerical data, taking AT as the symbolic feature subset of the l-th graph convolutional layer, the indiscernibility relation is shown in Formula (4):

[0086]

[0087] where, x(j)) ≤ α is a metric function with respect to AT’, and α > 0 is a threshold. Since the Euclidean distance is widely used, the Euclidean distance is taken as the metric function in this embodiment:

[0088] For incomplete data, taking AT as the symbolic feature subset of the l-th graph convolutional layer, the indiscernibility relation is shown in Formula (5):

[0089]

[0090] Based on the obtained feature weights and Formulas (3), (4), and (5), for the learners x(i) and x(j) represented as nodes in the graph, the calculation method of the weight of the edge between them is shown in Formula (6):

[0091]

[0092] where, τ is a threshold, and (x(i), x(j)) is the index function of x(i) and x(j) under the feature set Acc z Symbolic represents the symbolic data type, numerical represents the numerical data type, and incomplete represents the incomplete data type.

[0093] S13. Input the learner relationship graph into the learner anomaly detection model to predict the learning anomaly detection result of the learner.

[0094] Such as Figure 3As shown in the figure, in this embodiment, a granularity - adaptive graph wavelet convolutional neural network (GA - GWNN) that includes three processes: graph aggregation, graph restoration, and output is used as the learner anomaly detection model of this embodiment, effectively capturing different granularities of node features and graph topology information, thereby improving the efficiency of anomaly feature detection. Among them, the graph wavelet neural network replaces the Fourier transform in the traditional graph convolution with a wavelet transform, transforming the data into the frequency domain, making the calculation more efficient. In this embodiment, the construction process of the learner anomaly detection model includes: constructing an aggregation layer using a modularity - based graph algorithm, constructing a restoration layer using a skip - connection algorithm, and then constructing an output layer based on the aggregation layer and the restoration layer.

[0095] Specifically, after inputting the learner relationship graph into the learner anomaly detection model, as Figure 4 shown, the learner anomaly detection model performs the following steps:

[0096] S31. Partition the feature nodes in the learner relationship graph through the aggregation layer.

[0097] In this embodiment, the aggregation layer construction stage is mainly used to partition the learning feature nodes. The graph aggregation consists of repeating two operations: graph wavelet convolution and graph aggregation. Assume that the current layer is the l - th layer (l = 0, 1, ···, n Laggregation - 1), then the input of this module is the graph G of the l - th layer l , the graph representation is X l , the graph output is the graph G of the (l + 1) - th layer l+1 , the hidden nodes are represented as H l+1 , and the graph representation is X l+1 . To generate aggregation nodes from the coarse graph, this embodiment proposes a graph wavelet convolution mechanism based on the improved Louvain algorithm, which can obtain the optimal solution of the learner data topology graph and is used to solve the anomaly detection problem of online learning. Based on the above process, the aggregation steps of this embodiment are as follows:

[0098] First, predict the graph topology and graph representation based on the learner relationship graph, and at the same time use graph wavelet convolution to extract hidden feature nodes; then define all the feature nodes in the learner relationship graph as a community respectively, and initialize the community as an unlabeled state, where the feature nodes include hidden feature nodes and unhidden feature nodes; calculate the modularity of the preset community, and adjust the subordination relationship between the unlabeled community and the preset community according to the modularity, and mark the nodes corresponding to the community with the determined subordination relationship; then obtain the unlabeled nodes, calculate the normalized weights of the edges of all the first-order neighbor nodes of the unlabeled nodes, put the unlabeled nodes into the community with the highest normalized weight, and after marking the unlabeled nodes, construct an aggregation graph according to the marked nodes.

[0099] In this embodiment, adjusting the subordination relationship between the unlabeled community and the preset community and marking the nodes corresponding to the community with the determined subordination relationship can be achieved through the following methods:

[0100] Determine the preset difference of the modularity for moving the unlabeled community to the preset community; determine that the preset difference is greater than zero and the number of nodes in the neighbor community corresponding to the preset community is less than the preset number, and move the unlabeled community to the preset community; determine that the modularity is in a stable state and stop the moving process of the unlabeled community. Wherein, the preset number is the upper limit value of the number of nodes in the neighbor community of the current node.

[0101] Specifically, at layer l, according to graph G l obtain the adjacency matrix A l and the graph representation X l , and use graph wavelet convolution to extract the hidden node embedding H l+1 , which can be described as shown in formulas (7) and (8):

[0102] H l =X l W Formula (7)

[0103]

[0104] Among them, W is the trainable weight matrix of feature transformation, E l is the diagonal matrix graph convolution, ReLU is the activation function, and η w is the wavelet basis in the graph wavelet convolution network, w is the control parameter, and has the relationship shown in formula (9):

[0105] η w =UD w U T Formula (9)

[0106] Among them, U=(u1,u2,...,u n) is the orthogonal eigenvector corresponding to the Laplacian matrix of a graph, D w = diag(d(wμ1), d(wμ2),..., d(wμ n )) is its control matrix and where μ i is the i-th eigenvector of matrix D w . In this embodiment, the graph wavelet transform and the inverse wavelet are respectively defined as and x = η w x'.

[0107] Each learning feature node is defined as a community, and these communities are initialized as unlabeled. Calculate the modularity Q of the entire community. For each unlabeled node, the modularity gain for moving it to community C is a preset difference ΔQ. If ΔQ > 0 and the number of nodes in the neighbor community of the current node does not reach the upper limit, then add this node to the community and mark it until the modularity Q of the entire community no longer increases. The calculation process of the modularity is shown in formula (10):

[0108]

[0109] where m is the sum of the weights of all edges in graph G', G' represents a subgraph of G, N c is the number of communities, C i is the i-th community, is the edge weight of the i-th community, edgeC i is the edge connected to the vertices in community C i , is the edge weight of the edge connected to the vertices in community C i .

[0110] In addition, the modularity gain for moving node j to community C is calculated as shown in formula (11):

[0111]

[0112] where w1 is the sum of the weights of all edges from i to this node in graph G', and w2 is the sum of the weights of all edges of node i.

[0113] For the still unlabeled node x(i), calculate the normalized weight of its edges to all first-order neighbor nodes x(j), and put x(i) into the community to which x(j) with the highest normalized weight belongs and mark it. The process is expressed as shown in formula (12):

[0114]

[0115] where W edgeDenote the weight of the relational edge, W node Denote the weight of the node.

[0116] After the above operations, start to construct the aggregated graph G l+1 . First, define a matching matrix M l for passing information from graph G l+1 to G l,l+1 , including its structure, and its size is At layer l, when the node x(i) belongs to the super node x(i) in G l+1 , set the value of the i-th row and j-th column of M l,l+1 to 1. Among them, the adjacency matrix A l+1 of the aggregated graph G l+1 and the graph representation X l+1 are as shown in formulas (13) and (14):

[0117] A l+1 =(M l，l+1 )A l (M l，l+1 ) Formula (13)

[0118] X l+1 =(M l，l+1 )H l+1 Formula (14)

[0119] Graph G l+1 and its graph representation X l+1 are passed as the input of the next layer to the next layer, and the aggregation process from layer l to layers l + 1 and l + 2 can be vividly represented as Figure 5 shown.

[0120] S32. Adjust the partitioning process of the aggregation layer through the restoration layer.

[0121] In this embodiment, in order to minimize the information loss in the process of obtaining the final result from the aggregation module, several network layers for graph restoration are connected to their corresponding graph convolutional layers participating in the aggregation process. Specifically, the feature transformation of the graph wavelet convolutional network and graph convolution are used to generate the hidden feature nodes in the input graph of the restoration layer, and according to the hidden feature nodes in the input graph of the restoration layer, the graph representation of the input graph is restored by using the matching matrix. At the same time, skip connections are used to combine the hidden feature nodes extracted by the graph wavelet convolution and the hidden feature nodes in the input graph of the restoration layer. For example, the input of this module is the graph L at the n -l - 1 layer graph representation L and the output is the graph hidden node representation graph representation Then, during the processing of this embodiment, the feature transformation and graph convolution of the graph wavelet convolution network are also used to generate a graph of hidden node embeddings The matching matrix M is used to restore the graph representation. The aggregation process and the restoration process of the graph are symmetric. G l and have the same graph topology, so there is In addition, this embodiment also uses skip connections to connect H at layer l l+1 and the corresponding hidden node embeddings generated during the refinement process in the winding process Combined. Specifically, the calculation of the graph restoration process is shown in formula (15):

[0122]

[0123] n L -l-1 layer to n L -l, n L -l+1 layer of the restoration process is as Figure 6 shown

[0124] S33. Predict the label probabilities of the unlabeled nodes in the learner relationship graph through the classifier of the output layer

[0125] In this embodiment, based on the interaction between the above aggregation layer and restoration layer, a graph representation form is obtained during the graph reduction process of the output layer and finally the label prediction of the unlabeled nodes is obtained through the classifier. Specifically, at the output layer of the nth L layer of the GA-GWNN network, finally on use the graph wavelet convolution network with a Softmax classifier to output the probabilities of the unlabeled nodes, as shown in formulas (16), (17), (18) and (19):

[0126]

[0127]

[0128]

[0129]

[0130] Among them, is the trained N ano class weight matrix, SF is the prediction possibility of the N ano dimensional probability simplex, N ano is the number of learner anomaly types

[0131] In this embodiment, a loss function Γ as shown in formula (20) is defined to minimize the cross-entropy error of all labeled nodes V in the graph: label The cross-entropy error is minimized:

[0132]

[0133] where Y is is an indicator function, which has a value of 1 if and only if the label of node x(i) is s, and Sf is is the predicted probability that node x(i) belongs to class c.

[0134] In some embodiments, each module involved is differentiable, and the entire learner anomaly detection model formed is also differentiable. Therefore, this embodiment uses a deep learning optimization algorithm to globally train the learner anomaly detection model.

[0135] In addition, an embodiment of the present invention also discloses a computer program product or a computer program. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device can read the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to execute Figure 1 the method shown.

[0136] The above has described the embodiments of the present invention in detail with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art in the relevant technical field, various changes can be made without departing from the gist of the present invention. In addition, the embodiments of the present invention and the features in the embodiments can be combined with each other without conflict.

Claims

1. An anomaly learner detection method based on graph aggregation and recovery supported by granular computing, characterized in that Including the following steps: Obtain the learner data of the online learning platform; Perform data granulation on the learner data to obtain a learner relationship graph; Input the learner relationship graph into the learner anomaly detection model to predict the learning anomaly detection result of the learner; Among them, the learner anomaly detection model is constructed through the following steps: Construct an aggregation layer using a graph algorithm based on modularity; Construct a recovery layer using a skip connection algorithm; Construct an output layer based on the aggregation layer and the recovery layer; The performing data granulation on the learner data to obtain a learner relationship graph includes: Extract features related to learner anomalies from the learner data and calculate the weights of the features related to learner anomalies; Generate a learner relationship graph according to the features related to learner anomalies and the weights; The extracting features related to learner anomalies from the learner data and calculating the weights of the features related to learner anomalies includes: Extract features related to learner anomalies from the learner data to obtain a learner feature set, a feature set to be determined for importance, and a learner anomaly situation set; Substitute the learner feature set, the feature set to be determined for importance, and the learner anomaly situation set into rough set theory for calculation to obtain the first positive domain of the learner anomaly situation set under the learner feature set and the second positive domain of the learner anomaly situation set under the feature set to be determined for importance; Determine the importance definition of the feature set to be determined for importance with respect to the learner anomaly situation set according to the first positive domain and the second positive domain; Screen the feature set to be determined for importance according to the importance definition; Calculate the weight of each feature in the screened feature set to be determined for importance; The generating a learner relationship graph according to the features related to learner anomalies and the weights includes: Granulate the features related to learner anomalies according to the types of the features related to learner anomalies; Calculate the weight of the edge of each node in the learner relationship graph according to the granulation result and the weight of each feature; Generate a learner relationship graph according to the features related to learner anomalies and the weight of the edge of each node.

2. The anomaly learner detection method based on graph aggregation and recovery under the support of granular computing according to claim 1, wherein After the learner relationship graph is input into the learner anomaly detection model, the learner anomaly detection model performs the following steps: Partition the feature nodes in the learner relationship graph through the aggregation layer; Adjust the partitioning process of the aggregation layer through the recovery layer; Predict the label probability of the unlabeled nodes in the learner relationship graph through the classifier of the output layer.

3. An abnormal learner detection method based on graph aggregation and restoration under the support of granular computing according to claim 2, characterized in that, The partitioning the feature nodes in the learner relationship graph through the aggregation layer includes: Predict the graph topology and graph representation according to the learner relationship graph, and extract hidden feature nodes using graph wavelet convolution; Define all the feature nodes in the learner relationship graph as a community, and initialize the community as an unlabeled state. The feature nodes include hidden feature nodes and unhidden feature nodes; Calculate the modularity of the preset community; Adjust the subordination relationship between the unlabeled community and the preset community according to the modularity, and label the nodes corresponding to the community with the determined subordination relationship; Obtain the unlabeled nodes, and calculate the normalized weights of the edges of all first-order neighbor nodes of the unlabeled nodes; Put the unlabeled nodes into the community to which the highest normalized weight belongs, and label the unlabeled nodes; Construct an aggregation graph according to the labeled nodes.

4. An abnormal learner detection method based on graph aggregation and restoration under the support of granular computing according to claim 3, characterized in that The adjusting the subordination relationship between the unlabeled community and the preset community according to the modularity, and labeling the nodes corresponding to the community with the determined subordination relationship includes: Determine the preset difference in modularity for moving the unlabeled community to the preset community; Determine that the preset difference is greater than zero and the number of nodes in the neighbor community of the nodes corresponding to the preset community is less than the preset number, and move the unlabeled community to the preset community; Determine that the modularity is in a stable state, and stop the moving process of the unlabeled community.

5. An anomaly learner detection method based on graph aggregation and restoration under the support of granular computing according to claim 3, characterized in that The adjusting the partitioning process of the aggregation layer by the restoration layer includes: Use the feature transformation of the graph wavelet convolution network and graph convolution to generate the hidden feature nodes in the input graph of the restoration layer; According to the hidden feature nodes in the input graph of the restoration layer, use the matching matrix to restore the graph representation of the input graph.

6. The anomaly learner detection method based on graph aggregation and recovery under the support of granular computing according to claim 5, characterized in that, The adjusting the partitioning process of the aggregation layer by the restoration layer further includes: Use skip connections to combine the hidden feature nodes extracted by graph wavelet convolution and the hidden feature nodes in the input graph of the restoration layer.

7. The anomaly learner detection method based on graph aggregation and recovery supported by granular computing according to claim 1, wherein The learner anomaly detection model is a pre-trained model, and the training process includes: Use the deep learning optimization algorithm to globally train the learner anomaly detection model.

Images