Man-machine interaction flow prediction method based on collaborative graph encoder

Abstract

The invention relates to the technical field of man-machine interaction flow prediction in a public network, and discloses a man-machine interaction flow prediction method based on a collaborative graph encoder, which comprises the following steps: S1, collecting weblog data; s2, preprocessing the data; s3, constructing an isomorphic hypergraph data structure; s4, developing a hypergraph isomorphic auto-encoder model; s5, predicting the network traffic; by setting the step S3 and the step S4, a two-channel mechanism is established to collaboratively encode network data by adopting a collaboration graph encoder technology, and network flow data is predicted, so that the problems of poor detection effect, difficulty in modeling, poor time sequence and the like in a traditional method are avoided; the method can effectively improve the prediction of the network flow in the man-machine process, improves the protection capability of the network, and better retains the global feature information of the network through combining with the common graph convolution network modeling.

Description

Technical Field

[0001] The present invention relates to the technical field of human-computer interaction traffic prediction in a public network, and more particularly to a human-computer interaction traffic prediction method based on a collaborative graph encoder. Background Art

[0002] Human-computer interaction traffic prediction is an important research direction in network planning and management, mainly to predict the network traffic generated in future human-computer interactions by analyzing network channel traffic, access information, etc. With the rapid development of Internet technology, the network distribution tends to be centralized, distributed, and complex. Especially in the process of human-computer interaction, network traffic prediction has the characteristics of high real-time requirements, fast changes, and difficulty in prediction, making it difficult to effectively predict the correlation between network planning and network traffic in the process of human-computer interaction. Therefore, how to effectively predict network traffic is the key issue in network flow prediction during human-computer interaction.

[0003] The existing network traffic prediction methods mainly include the following: 1) Simple data analysis: Based on the collected historical network data, obvious network features such as the number of IP accesses and the number of accessed bytes are detected, and at the same time, known human-computer interaction network planning patterns are identified; 2) Pattern recognition: By collecting network log data during the interaction process, a prediction model is established using historical log data, and traditional machine learning algorithms such as decision trees and GBDT are used to predict traffic in different network patterns. However, this method is difficult to effectively model the network structure, thus affecting the effect of network traffic prediction; 3) Deep learning: Using deep learning models, especially graph neural networks, can better model network relationships, process complex data structures, and improve the accuracy and efficiency of human-computer interaction network traffic prediction. However, the existing network traffic prediction methods have some limitations and deficiencies: 1) Single data feature: Most methods rely on a single type of data source such as historical network log data, which limits the comprehensiveness of the data; 2) Model limitation: The prediction model cannot consider the temporal characteristics of different network environments changing over time, resulting in inaccurate model prediction; 3) High false alarm rate: There is a large amount of noise in the data of the interaction network, and traditional methods often produce a high false alarm rate, bringing interference to network planning. Summary of the Invention

[0004] To overcome the above-mentioned defects of the prior art, the present invention provides a human-computer interaction traffic prediction method based on a collaborative graph encoder, which is applicable to predicting network flow in different stages of complex human-computer interaction processes, facilitating network planning and management. A hypergraph structure is adopted to express the complex relationships between vertices in network traffic prediction, such as the relationship between IP requests and network links: one access starting point can correspond to multiple networks; and a dual-channel deep learning model based on the hypergraph structure is designed to effectively improve the traffic prediction effect under complex networks, while combining ordinary graph convolutional network modeling to better retain the global feature information of the network; to solve the problems existing in the above-mentioned background technology.

[0005] The present invention provides the following technical solutions: A human-computer interaction traffic prediction method based on a collaborative graph encoder, comprising the following steps: Step S1: Network log data collection: Collect network log data based on NetFlow; Step S2: Data preprocessing: Process the data into directly usable data; Step S3: Construct an isomorphic hypergraph data structure: Construct a hypergraph neural network according to the data divided in Step S2 to form an isomorphic hypergraph data structure; Step S4: Develop a hypergraph isomorphic autoencoder model: Based on the isomorphic hypergraph data structure constructed in Step S3, add a self-attention mechanism based on the central node to obtain an improved hypergraph neural network; Construct a hypergraph isomorphic autoencoder based on the improved hypergraph neural network; Step S5: Network traffic prediction: Take the network log data features within each five-minute interval as a training sample, obtain the feature vectors of hypergraph vertices and link hyperedges through the hypergraph autoencoder, and then put all the feature vectors into a set of fully connected network layers to output the predicted network traffic for the next moment.

[0006] Preferably, the network log data is access log data generated in the network, mainly including the detailed information of IP sessions such as the source address, destination address, port number of the accessed IP, and the number of bytes transmitted.

[0007] Preferably, in Step S2, the data is divided according to the time series, and the time slice division interval is 5 minutes, that is, the network log data is sliced at intervals of 5-minute time intervals, and each 5-minute interval will contain n access IP information.

[0008] Preferably, each accessed IP entity in the hypergraph neural network is represented as a vertex in the hypergraph, and the time series chain of IP access is recorded as a hyperedge of the hypergraph; Construct a hypergraph vertex set to initialize the encoding features of each vertex of the hypergraph, construct a hyperedge set, and customize the weight calculation method of the hyperedge to assign a weight value to the hyperedge.

[0009] Preferably, the hypergraph autoencoder includes an encoder and a decoder, and the core part is an improved hypergraph neural network; through the message passing mechanism and the aggregation function, the hypergraph vertex embeddings and hyperedge features are updated. After training, the low-dimensional dense hidden layer vector of the hypergraph autoencoder is the required hypergraph vertex feature vector, and the feature vector of each hyperedge is obtained through the aggregation operation.

[0010] Preferably, the preprocessing steps in step S2 include removing duplicate log data, removing null data, normalizing the log data, and outlier processing; The formula for the normalization process is expressed as: (1); where, represents the data after normalization processing, represents the data that needs to be normalized, represents the average value of all data, represents the standard deviation of all data; The specific method of the outlier processing is as follows: If there is one or more feature values in the feature data that are outlier points located at the extreme values, then the original feature data sample is directly deleted, that is, this feature data is directly deleted; If the feature value in the feature data is an outlier point at a non-extreme value, then the median of the normal data in the feature data is used to replace the outlier data, that is, the median of the feature data is used for correction.

[0011] Preferably, the hypergraph data structure in step S3 generally uses a triple to represent, where, represents the hypergraph vertex set, and each degree network access IP entity corresponds to a vertex in, is defined as the hyperedge set, represents the hyperedge weight matrix; each hyperedge of the hypergraph can connect multiple vertices, and the association relationship between the hyperedge and the vertex can be represented by an incidence matrix to represent, each element in is defined as follows: (2); where, represents the th hyperedge of the hypergraph, represents the th connected vertex; represents the element formed by the th hyperedge and the th connected vertex.

[0012] Preferably, the The degree of is expressed by the formula: (3); where, represents the degree of, represents the weight corresponding to the hyperedge; represents the vertices in the hypergraph; The said hyperedge The degree of is expressed by the formula: (4); where, represents the hyperedge in the hypergraph; represents the degree of the hyperedge; Customize the hyperedge weight, and the formula is expressed as: (5); where, represents the Euclidean distance between vertices, represents the distance decay coefficient, represents the th vertex; represents the weight of. Preferably, in the step S4 of developing the hypergraph isomorphism autoencoder model, the isomorphic graph means that the vertex types in the graph are the same. During the process of developing the hypergraph isomorphism autoencoder model, hypergraph convolution is required, and the formula is expressed as: (7); where, represents the - layer graph convolution network; represents the non - linear activation function; represents the degree matrix of vertices; represents the - layer graph neural convolution layer; represents the weight matrix of the - layer network training; Apply the attention mechanism of the central vertex to the hypergraph convolution process, and its calculation formula is as follows: (8); where, represents the non - linear activation function; and represent two vertices on the same hyperedge; is used to calculate the correlation coefficient between hypergraph vertices; represents the vertex and whether they are connected; N represents the number of vertices.

[0013] Preferably, a residual module in the neural network is added during the hypergraph convolution process, then formula (7) becomes: (9); For the hypergraph autoencoder, the above hypergraph convolutional network is selected as the encoder module and the decoder module; for the encoder network, two parameters and are obtained through the following two hypergraph convolutional networks: (10); (11); where represents the feature matrix; represents the vertex adjacency matrix; The decoder network determines whether there is an edge between nodes through the inner product of the hidden vectors.

[0014] Technical effects and advantages of the present invention: By providing step S3 and step S4, the present invention is beneficial to establishing a dual-channel mechanism to collaboratively encode network data and predict network traffic data by adopting the collaborative graph encoder technology, avoiding problems such as poor detection effect, difficult modeling, and poor timeliness in traditional methods; it can effectively improve the prediction of network traffic in the human-machine process and improve the network protection ability; it is applicable to predicting network flow in different stages of complex human-machine interaction processes, facilitating network planning and management, using a hypergraph structure to express the complex relationships between vertices in network traffic prediction, such as the relationship between IP requests and network links: a single access starting point can correspond to multiple networks; and a dual-channel deep learning model based on the hypergraph structure is designed to effectively improve the traffic prediction effect under complex networks, and at the same time, combined with the ordinary graph convolutional network modeling, better retain the global feature information of the network. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flowchart of the human-machine interaction traffic prediction method based on the collaborative graph encoder of the present invention.

[0016] Figure 2 is a flowchart of the human-machine interaction traffic prediction based on the hypergraph isomorphism autoencoder of the present invention.

[0017] Figure 3 is a schematic diagram of the hypergraph structure of the present invention.

[0018] Figure 4 is a flowchart of the log data collection and analysis of the present invention.

[0019] Figure 5 is a schematic diagram of the improved hypergraph neural network of the present invention.

[0020] Figure 6 is a framework diagram of the hypergraph autoencoder of the present invention.

[0021] Figure 7Schematic diagram of the fully connected network of the present invention. Detailed implementation manners

[0022] Next, the technical solutions of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the present invention. In addition, the forms of the respective structures described in the following embodiments are merely examples, and a human-computer interaction traffic prediction method based on a collaborative graph encoder involved in the present invention is not limited to the respective structures described in the following embodiments. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present invention.

[0023] As Figures 1 - 7 shown, the present invention provides a human-computer interaction traffic prediction method based on a collaborative graph encoder, including the following steps: Step S1: Network log data collection: Collect network log data based on NetFlow. The network log data is access log data generated in the network, mainly including detailed information of IP sessions such as the source address, destination address, port number of the access IP, and the number of bytes transmitted. NetFlow statistically records network session information through flow table entries and provides a user interface for viewing real-time data collection information. When data is transmitted in the network, the NetFlow device classifies the data packets transmitted in the network. NetFlow creates a flow table entry, and the flow table entry records the detailed information of the session flow. When the data packet matches an existing flow table entry successfully, the counter of the corresponding entry is updated. When the number of the counters reaches a threshold, the flow table is packaged into a NetFlow data packet and sent to the NetFlow collector for processing and analysis; Step S2: Data preprocessing: Process the data into directly usable data; divide the data according to the time series, and the time slice division interval is 5 minutes, that is, the network log data is sliced at intervals of 5 minutes. Each 5-minute interval contains n access IP information. After obtaining the data, first perform duplicate removal processing on the data. Due to the delay in collection, duplicate data may be collected. Then, perform deletion or filling operations on the null value data. Finally, to ensure that the data features are in the same dimension, perform normalization processing on the data; Step S3: Construct an isomorphic hypergraph data structure: Construct a hypergraph neural network based on the data partitioned in Step S2 to form an isomorphic hypergraph data structure; each accessed IP entity in the hypergraph neural network is represented as a vertex in the hypergraph, and the time series chain of IP accesses is recorded as a hyperedge of the hypergraph; construct a hypergraph vertex set to initialize the encoding features for each vertex of the hypergraph, construct a hyperedge set, use an adjacency matrix to describe the connection relationship between hyperedges and vertices, customize the weight calculation method for hyperedges, and assign weight values to hyperedges; since the constructed hypergraph data structure is an isomorphic graph, that is, the properties of network nodes are the same; Step S4: Develop a hypergraph isomorphic autoencoder model: Based on the isomorphic hypergraph data structure constructed in Step S3, add a self-attention mechanism based on the central node to obtain an improved hypergraph neural network; construct a hypergraph isomorphic autoencoder based on the improved hypergraph neural network; the hypergraph isomorphic autoencoder includes an encoder and a decoder, and the core part is the improved hypergraph neural network; update the hypergraph vertex embeddings and hyperedge features through a message passing mechanism and an aggregation function. After training, the low-dimensional dense hidden layer vector of the hypergraph autoencoder is the required hypergraph vertex feature vector, and the feature vector of each hyperedge is obtained through an aggregation operation; Step S5: Network traffic prediction: Use the network log data features within each five-minute interval as a training sample, obtain the feature vectors of hypergraph vertices and link hyperedges through the hypergraph autoencoder, and then put all the feature vectors into a set of fully connected network layers to output the predicted network traffic that will occur at a certain stage in the next moment.

[0024] In this embodiment, it should be specifically noted that NetFlow is a network data stream collection framework developed by Cisco and can be used to monitor and analyze network traffic data; first, enable the NetFlow function on the router that supports NetFlow, configure NetFlow to the newer version v9, and configure the sampling rate according to the actual network traffic. If the network traffic is too large, the sampling rate can be used to reduce the number of generated technologies and reduce the sampling pressure of NetFlow. Initially, set the sampling rate to 100 packets / time; second, specify the destination of NetFlow export, and the address of the NetFLow collector to which it is to be sent needs to be specified. To increase the transmission efficiency, configure the transmission protocol as UDP and the corresponding port number. According to the corresponding NetFlow version, define the data fields to be collected, such as the source IP address, destination IP address, transmitted byte count, etc. The specific data fields collected are shown in the following table: Finally, the NetFlow collector needs to be installed, and the data of the NetFlow Collector is passed into the data analysis module for the next data cleaning work.

[0025] In this embodiment, it should be specifically noted that the preprocessing steps in step S2 include, but are not limited to, removing duplicate log data, removing null data, normalizing log data, and outlier processing, etc.; null data and duplicate data will affect the prediction accuracy of the model. The methods in Python can be used to directly remove duplicate data and null data from the original dataset; in order to avoid a certain feature having too much influence on the model, the data is further normalized, and at the same time, the training efficiency of the model can be improved; the formula for the normalization process is expressed as: (1); where represents the data after normalization processing, represents the data that needs to be normalized, represents the average value of all data, represents the standard deviation of all data; The purpose of the outlier processing is to remove the possible noise data in the network log data and avoid affecting the stability of the model; for numerical features, first, the average value and median of all numerical feature data need to be calculated. According to the capping principle, if the distance between the data sample and the average value is greater than outside, it is considered an outlier; the specific method of the outlier processing is: If there is one or more feature values in the feature data that are outliers at the extreme values, such as the occupancy rate, flight number, etc. in a certain feature data are at the extreme value outliers, then the original feature data sample is directly deleted, that is, this feature data is directly deleted; If the feature value in the feature data is an outlier at a non-extreme value, then the median of the normal data in the feature data is used to replace the outlier data, that is, it is corrected using the median of the feature data; Further, for categorical vectors such as device codes, LabelEncoder encoding is used to convert them into numerical features for the convenience of model learning.

[0026] In this embodiment, it should be specifically noted that the hypergraph data structure in step S3 generally uses a triple to represent, where represents the hypergraph vertex set, and each degree network access IP entity corresponds to a vertex in, is defined as the hyperedge set, represents the hyperedge weight matrix; each hyperedge of the hypergraph can connect multiple vertices. For the association relationship between the hyperedge and the vertex, it can be represented by an incidence matrix to represent, Each element in is defined as follows: (2); where represents the th hyperedge of the hypergraph, represents the th vertex of the connection; represents the th hyperedge and the th vertex of the connection, forming an element; The degree is expressed by the formula: (3); where represents the degree, represents the weight corresponding to the hyperedge; represents the vertices in the hypergraph; The degree of the hyperedge is expressed by the formula: (4); where represents the hyperedge in the hypergraph; represents the degree of the hyperedge; The custom definition of the hyperedge weight is expressed by the formula: (5); where represents the Euclidean distance between vertices, represents the distance attenuation coefficient, represents the th vertex; represents the weight; Equation (2) defines the basic structure of the hypergraph, and equations (3) and (4) are used to calculate the hypergraph Laplacian matrix: (6); where represents the hypergraph Laplacian matrix; represents the vertex degree matrix; represents the hyperedge degree matrix; The Laplacian matrix can describe more global and in-depth hypergraph structure information; equation (5) defines the calculation method of the hyperedge weight in the hypergraph structure.

[0027] In this embodiment, it should be specifically noted that in the step S4 of developing the hypergraph isomorphism autoencoder model, the isomorphic graph means that the vertex types in the graph are the same. For example, the graph structure nodes in this embodiment are all network IP access entities, so it is an isomorphic graph; during the development of the hypergraph isomorphism autoencoder model, hypergraph convolution needs to be performed, and the formula is expressed as: (7); where represents the layer graph convolutional network; represents the non-linear activation function; Denote the degree matrix of vertices; Denote layer graph neural convolution layer; Denote weight matrix of layer network training; Since the correlation of neighbor vertices is different with respect to a certain central vertex, this embodiment proposes an attention mechanism based on the central vertex and applies it to the hypergraph convolution process to improve the vertex feature aggregation ability. Its calculation formula is as follows: (8); where Denote the non-linear activation function; and Denote two vertices on the same hyperedge; Used to calculate the correlation coefficient between hypergraph vertices; Denote vertex and Whether they are connected; N denotes the number of vertices; In order to retain the basic information of the original vertex features during the convolution process, this embodiment adds a residual module in the neural network, and formula (7) becomes: (9); For the hypergraph autoencoder, select the above hypergraph convolutional network as the encoder module and decoder module; for the encoder network, first obtain two parameters and , obtained through the following two hypergraph convolutional networks: (10); (11); where Denote the feature data; Denote the vertex adjacency matrix; The decoder network determines whether there is an edge between nodes through the inner product of hidden vectors.

[0028] In this embodiment, it should be specifically noted that the hypergraph isomorphism autoencoder can map the original feature data into a low-dimensional feature vector through the hidden layer. The decoder model also uses the hypergraph convolution model, and then multiplies the obtained features to judge the possibility of association between vertices. The general convolution mode of the encoder and decoder is as follows: (12); where Denote the encoded feature vector; Denote the activation function; Denote the training weight matrix; Denote the bias matrix; The entire process of the hypergraph autoencoder is trained in an unsupervised manner, and the objective of the training loss function is to minimize the structural difference between the reconstructed hypergraph and the original hypergraph. In this embodiment, a multi-encoder stacking method will be used, that is, the output result of one encoder is input into the next encoder to obtain more high-order low-dimensional vertex embedding information. For the modeling of the ordinary graph convolution channel, referring to formula (9) is sufficient, and only the hypergraph structure needs to be replaced with the graph structure.

[0029] In this embodiment, it should be specifically noted that after obtaining the result of the hypergraph autoencoder in the network traffic prediction of step S5, a fully connected layer network is established , converting the unsupervised hypergraph autoencoding problem into a classification problem. The fully connected layer contains a weight matrix and a bias vector . These parameters need to be learned during the training process of the fully connected network. At the same time, in order to introduce non-linearity, an activation function RELU also needs to be added. Finally, the prediction of the human-computer interaction process on the network in a certain network mode is provided, which has guiding significance for network planning.

[0030] Finally, the above are only the preferred embodiments of the present invention and are not used to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0031] The above is only the specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, and all of them should be covered by the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A human-computer interaction traffic prediction method based on a collaborative graph encoder, characterized in that: It includes the following steps: Step S1: Network log data collection: Collect network log data based on NetFlow; Step S2: Data preprocessing: Process the data into directly usable data; Step S3: Constructing an isomorphic hypergraph data structure: Construct a hypergraph neural network according to the data divided in Step S2 to form an isomorphic hypergraph data structure; Step S4: Developing a hypergraph isomorphic autoencoder model: Based on the isomorphic hypergraph data structure constructed in Step S3, add a self-attention mechanism based on the central node to obtain an improved hypergraph neural network; Construct a hypergraph isomorphic autoencoder based on the improved hypergraph neural network; Step S5: Network traffic prediction: Use the network log data features within each five-minute interval as a training sample, obtain the feature vectors of hypergraph vertices and link hyperedges through the hypergraph autoencoder, and then put all the feature vectors into a set of fully connected network layers to output the predicted network traffic for the next moment.

2. The human-computer interaction traffic prediction method based on a collaborative graph encoder according to claim 1, wherein: The network log data is the access log data generated in the network, mainly including the detailed information of the IP session such as the source address, destination address, port number of the access IP, and the number of bytes transmitted.

3. The human-computer interaction traffic prediction method based on a collaborative graph encoder according to claim 2, wherein: In Step S2, the data is divided according to the time series, and the time slice division interval is 5 minutes, that is, the network log data is sliced at intervals of 5 minutes, and each 5-minute interval will contain n access IP information.

4. The human-computer interaction traffic prediction method based on a collaborative graph encoder according to claim 3, wherein: In the hypergraph neural network, each accessed IP entity is represented as a vertex in the hypergraph, and the time series chain of IP access is recorded as a hyperedge of the hypergraph; Construct a hypergraph vertex set to initialize the encoding features for each vertex of the hypergraph, construct a hyperedge set, and customize the weight calculation method of the hyperedge to assign a weight value to the hyperedge.

5. The human-computer interaction traffic prediction method based on a collaborative graph encoder according to claim 4, wherein: The hypergraph isomorphic autoencoder includes an encoder and a decoder, and the core part is the improved hypergraph neural network; Update the hypergraph vertex embedding and hyperedge features through the message passing mechanism and the aggregation function. After training, the low-dimensional dense hidden layer vector of the hypergraph autoencoder is the required hypergraph vertex feature vector, and the feature vector of each hyperedge is obtained through the aggregation operation.

6. The human-computer interaction traffic prediction method based on a collaborative graph encoder according to claim 5, wherein: The preprocessing steps in Step S2 include removing duplicate log data, removing null data, normalizing the log data, and outlier processing; The formula for the normalization process is expressed as: (1); where, represents the data after normalization, represents the data that needs to be normalized, represents the average value of all data, represents the standard deviation of all data; The specific method for outlier processing is: If there is one or more feature values in the feature data that are outlier points at the extreme values, directly delete the original feature data sample, that is, directly delete this feature data; If the feature value in the feature data is an outlier point at a non-extreme value, use the median of the normal data in the feature data to replace the outlier data, that is, correct it using the median of the feature data.

7. A human-computer interaction traffic prediction method based on a collaborative graph encoder according to claim 6, characterized in that: In step S3, the hypergraph data structure generally uses a triple to represent. Among them, represents the hypergraph vertex set, and each degree network access IP entity corresponds to a vertex in it. is defined as the hyperedge set. represents the hyperedge weight matrix; each hyperedge of the hypergraph can connect multiple vertices. For the association relationship between hyperedges and vertices, it can be represented by an incidence matrix to represent. Each element in is defined as follows: (2); where, represents the -th hyperedge of the hypergraph, represents the -th vertex of the connection; represents the element formed by the -th hyperedge and the -th vertex of the connection.

8. A human-computer interaction traffic prediction method based on a collaborative graph encoder according to claim 7, characterized in that: Said The degree is expressed by the formula: (3); among which, express The degree, represents the weight corresponding to the hyperedge; represents a vertex in a hypergraph; The hyperedge has its degree expressed by the formula: (4); where represents the hyperedge in the hypergraph; represents the degree of the hyperedge ; Customize the hyperedge weight, and the formula is expressed as: (5); where, represents the Euclidean distance between vertices, represents the distance attenuation coefficient, represents the -th vertex; represents 's weight.

9. The human-computer interaction traffic prediction method based on a collaborative graph encoder according to claim 8, wherein: In the process of developing the hypergraph isomorphic autoencoder model in Step S4, an isomorphic graph means that the vertex types in the graph are the same. During the development of the hypergraph isomorphic autoencoder model, hypergraph convolution is required, and the formula is expressed as: (7); among which, represents a layer of graph convolutional network; represents a non-linear activation function; represents the degree matrix of vertices; represents a layer of graph neural convolutional layer; represents the weight matrix of layer network training; Apply the attention mechanism of the central vertex to the hypergraph convolution process, and its calculation formula is as follows: (8); wherein, represents a non-linear activation function; and represent two vertices that are on the same hyperedge; is used to calculate the correlation coefficient between the vertices of the hypergraph; represents vertex and whether they are connected; N represents the number of vertices.

10. A human-computer interaction traffic prediction method based on a collaborative graph encoder according to claim 9, characterized in that: In the hypergraph convolution process, a residual module in the neural network is added, and formula (7) becomes: （9）； For the hypergraph autoencoder, the above hypergraph convolutional network is selected as the encoder module and the decoder module; for the encoder network, two parameters and are obtained through the following two hypergraph convolutional networks: (10); (11); wherein, represents feature data; represents the vertex adjacency matrix; The decoder network determines whether there is an edge between nodes through the inner product of hidden vectors.

Images