A data-driven communication network performance prediction method and apparatus

Abstract

The present invention discloses a data-driven communication network performance prediction method and apparatus. The method includes: obtaining a network information set; constructing network graph information based on the network information set; and using a trained network performance estimation model to process the network graph information to obtain a predicted result value of the network performance. Based on the network structure of the communication network, the present invention establishes a static network graph and a dynamic network graph, realizes a comprehensive description of the communication network result, and specifically establishes corresponding data variables and structures from perspectives such as data queues and flows, ensuring the accuracy of the evaluation result.

Description

Technical Field

[0001] The present invention relates to the fields of communications and artificial intelligence, and particularly to a data-driven communication network performance prediction method and apparatus. Background Art

[0002] Communication operators such as mobile networks and data center networks often need to predict the impact of various changes on network performance, such as new workloads, changes in topology, or network configurations. The most common method for predicting network performance is to use discrete-time simulators at the packet level, such as ns-3, omnet++, etc. However, the running efficiency of such network simulators is very low. Simulating a topology with 6,000 hosts takes 10 hours, while the simulation time is only 5 seconds. Any packet-level network simulator has inherent scalability limitations in high-speed, large-scale networks. As time goes by, this problem will become more serious as networks become larger and faster. Therefore, it is necessary to study more efficient methods to achieve accurate estimation of network performance. Summary of the Invention

[0003] The present invention mainly solves the problem of how to improve the efficiency of network simulation and fault prediction, and discloses a data-driven communication network performance prediction method and apparatus.

[0004] In a first aspect of an embodiment of the present invention, a data-driven communication network performance prediction method is disclosed, including:

[0005] S1, obtaining a network information set;

[0006] S2, constructing network graph information based on the network information set;

[0007] S3, using a trained network performance estimation model to process the network graph information to obtain a predicted result value of network performance.

[0008] The network information set includes communication network information at each moment; The communication network information at a moment is defined as , where the router information set is represented as , is the number of routers, represents the i-th router, the link information set is represented as , represents the i-th link information, represents the total number of links, the queue information set is represented as , represents the i-th queue information, is the total number of queues, and the routing policy set of the router at time t is represented as , is a router at time routing policy; The set of flow information at time is , represents the number of flows, represents the i-th flow information at time t; the flow information at time t is the set of queues and links through which the data sequence of the communication network at time t flows; the data sequence of the communication network at time t is determined according to the routing policy determined; the i-th flow information at time t is expressed as , and respectively represent the starting queue and starting link through which the flow passes, and respectively represent the terminating queue and terminating link through which the flow passes.

[0009] Based on the network information set, the network graph information is constructed, including:

[0010] S21, based on the network information set, a static influence graph set is constructed; the static influence graph set includes several static influence graphs;

[0011] The expression of the static influence graph set is , is the number of time steps, represents the static influence graph at the first time, represents the static influence graph at the T-th time, and each static influence graph includes a node set and a feature set ; the node set includes node information and feature information; the node information includes flow information, link information, and queue information; the feature information includes flow node features, link node features, and queue node features;

[0012] S22, using a node type mapping function, map the node information in the static influence graph to node type information;

[0013] S23, using a relationship type mapping function, map the feature information in the static influence graph to influence relationship information;

[0014] S24, using the node type information and influence relationship information, construct a dynamic influence graph;

[0015] S25, using the dynamic influence graph and the static influence graph set, construct the network graph information.

[0016] The node type mapping function The expression of is:

[0017] ,

[0018] Among them, represents node information, is the node type information, and the value range of the node type information is , which respectively represent the flow type, link type, and queue type;

[0019] The relationship type mapping function has the following expression:

[0020] ,

[0021] Among them, represents feature information, represents influence relationship information, and the value range of the influence relationship information is , is the link type 's influence relationship information on the flow type , is the queue type 's influence relationship information on the flow type .

[0022] The network performance estimation model includes a heterogeneous graph multi-head attention layer, a temporal multi-head attention layer, and a multi-metric output layer;

[0023] The heterogeneous graph multi-head attention layer is used to process the network graph information to obtain an output vector; the input end of the heterogeneous graph multi-head attention layer is used to receive the network graph information; the output end of the heterogeneous graph multi-head attention layer is connected to the input end of the temporal multi-head attention layer;

[0024] The temporal multi-head attention layer is used to process the output vector of the heterogeneous graph multi-head attention layer to obtain a node embedding feature vector; the output end of the temporal multi-head attention layer is connected to the input end of the multi-metric output layer;

[0025] The multi-metric output layer is used to process the node embedding feature vector to obtain a predicted result value of the network performance; the output end of the multi-metric output layer is used to output the predicted result value of the network performance; the predicted result value of the network performance includes predicted result values of the flow throughput, flow delay, link utilization rate, and queuing delay of the communication network.

[0026] The heterogeneous graph multi-head attention layer includes a first processing model, a second processing model, and a third processing model; the second processing model is respectively connected to the first processing model and the third processing model;

[0027] The calculation expression of the first processing model is as follows:

[0028] ,

[0029] wherein, is the node type information corresponding spatial transformation matrix, represents the output vector after the i-th node information of the static influence diagram set passes through the k-th feature mapping, represents the feature of the i-th node of the j-th static influence diagram in the static influence diagram set, , represents the node type information obtained by inputting the i-th node information of the static influence diagram set into the node type mapping function ;

[0030] The expression of the second processing model is:

[0031] ,

[0032] ,

[0033] wherein, represents the influence value of the j-th node of the static influence diagram on the i-th node through an edge of type r, represents the vector concatenation operation, represents the output vector after the j-th node information of the static influence diagram set passes through the k-th feature mapping, is the attention parameter vector of the edge of type , represents the position encoding value of the edge from the j-th node to the i-th node in the flow information, represents the -th node corresponding to the weight coefficient of the j-th node, represents the node connected to the neighbor nodes through an edge of type , represents the activation function;

[0034] The expression of the third processing model is:

[0035] ,

[0036] wherein, , is the type value of the node , , is the set of types of relevant edges of the node with node type , Represents the edge type variable, is the value range of, represents the output vector after the (k + 1)-th feature mapping of the information of the i-th node in the static influence graph set, represents the output vector after the k-th feature mapping of the information of the j-th node in the static influence graph set; head represents the attention head number, is the number of attention heads of the heterogeneous graph multi-head attention layer, is an injective function, implemented using a single-layer multi-layer perceptron, for mapping the information of different types of node influence relationships and multi-head attention information to the feature space of the same dimension.

[0037] The input sequence of the temporal multi-head attention layer is where is the output result of the heterogeneous graph multi-head attention layer at time t for the type value , and the length of the input sequence is the time window ;

[0038] The temporal multi-head attention layer includes a fourth processing model and a fifth processing model; the fourth processing model and the fifth processing model are connected;

[0039] The processing process of the fourth processing model includes:

[0040] Perform an update operation on the input sequence, and the expression of the update operation is:

[0041] ,

[0042] where is the encoding function for the value of the current time t, is the sin encoding function;

[0043] For each in perform the update operation to obtain the updated input sequence;

[0044] The processing process of the fifth processing model includes:

[0045] ,

[0046] ,

[0047] ,

[0048] ,

[0049] where , is the attention weight matrix determined by the scaled dot-product attention function, is the attention weights at time t and time p, is the time influence degree value between time t and time k, where k is the time value of the time window, , is the time influence degree value between time t and time p, is the vector dimension of the node embedding feature vector, type value The query matrix of type value The key matrix of type value The value matrix of , is the number of time attention heads in the temporal multi-head attention layer, is the preset mapping matrix, represents the node embedding feature vector obtained after fusing historical time features, and Concat represents the vector concatenation operation, represents the output result of the first temporal attention layer at time t for type value , and so on, represents the output result of the th temporal attention layer at time t for type value , represents the element in the t-th row and p-th column of matrix .

[0050] The position encoding value , is calculated by ; Since the link and the queue have different performance characteristics due to their different positions in the flow. Therefore, it is necessary to encode the position sequence numbers of the link and the queue in the flow . When the types of node and node are not both . In this work, the positions are encoded using common sine and cosine functions with different frequencies as follows

[0051] ,

[0052] ,

[0053] where is the sequence number of the position of the link or the queue in the flow, represents the th element in this position vector, is the floor of the quotient of the position subscript of the vector divided by 2. ​

[0054] The sin coding function can be implemented using the sin function, which represents the position of the current time t within the time window .

[0055] In the second aspect of the present invention, a data-driven communication network performance prediction device is disclosed. The device includes:

[0056] a memory storing executable program code;

[0057] a processor coupled to the memory;

[0058] The processor invokes the executable program code stored in the memory to execute the above-described data-driven communication network performance prediction method.

[0059] In the third aspect of the present invention, a computer-readable storage medium is disclosed. The computer-readable storage medium stores computer instructions that, when invoked by a computer, are used to execute the above-described data-driven communication network performance prediction method.

[0060] In the fourth aspect of the present invention, an information data processing terminal is disclosed. The information data processing terminal is used to implement the above-described data-driven communication network performance prediction method.

[0061] The beneficial effects of the present invention are as follows:

[0062] A data-driven communication network performance prediction method disclosed by the present invention first extracts key elements (flows, links, queues) in the communication network, and constructs an influence graph by extracting the influence relationships between them. Then, using the designed heterogeneous graph multi-head attention layer, the space between elements in the influence graph is extracted. On the basis of obtaining the spatial information, using the designed temporal multi-head self-attention mechanism, the time-dependent features of the influence graph are extracted to form the final embedding of the key element features. Finally, using the readout function, multi-granularity network performance estimation is achieved, providing support for communication network performance analysis and network optimization.

[0063] Based on the network structure of the communication network, the present invention establishes a static network graph and a dynamic network graph, achieving a comprehensive description of the communication network structure, and specifically establishing corresponding data variables and structures from the perspectives of data queues, flows, etc., ensuring the accuracy of the evaluation results.

[0064] By fusing spatial features and temporal features, the present invention achieves a comprehensive evaluation of the communication network performance, improving the comprehensiveness of performance prediction.

[0065] Other features and advantages of the present invention will be set forth in the following description, and in part will be obvious from the description, or may be learned by practice of the present invention. The objectives and other advantages of the present invention may be realized and attained by the structure particularly pointed out in the written description, claims, as well as the drawings. Description of the Drawings

[0066] Figure 1 is the implementation flowchart of the method of the present invention;

[0067] Figure 2 is the composition schematic diagram of the heterogeneous graph multi-head attention layer of the network performance estimation model of the present invention;

[0068] Figure 3 is the composition schematic diagram of the temporal multi-head attention layer of the network performance estimation model of the present invention;

[0069] Figure 4 is the composition schematic diagram of the multi-index output layer of the network performance estimation model of the present invention. Detailed Embodiment

[0070] To better understand the content of the present invention, an embodiment is given here.

[0071] Figure 1 is the implementation flowchart of the method of the present invention; Figure 2 is the composition schematic diagram of the heterogeneous graph multi-head attention layer of the network performance estimation model of the present invention; Figure 3 is the composition schematic diagram of the temporal multi-head attention layer of the network performance estimation model of the present invention; Figure 4 is the composition schematic diagram of the multi-index output layer of the network performance estimation model of the present invention.

[0072] In the first aspect of the embodiment of the present invention, a data-driven communication network performance prediction method is disclosed, including:

[0073] S1, obtaining a network information set;

[0074] S2, based on the network information set, constructing network graph information;

[0075] S3, using the trained network performance estimation model to process the network graph information to obtain a predicted result value of the network performance.

[0076] The network information set includes communication network information at each moment; The communication network information at the moment is defined as , where the router information set is denoted as , is the number of routers, Denote the i-th router, and the set of link information is denoted as , Denote the i-th link information, Denote the total number of links, and the set of queue information is denoted as , Denote the i-th queue information, is the total number of queues, and the set of routing policies of the router at time t is denoted as , is the router at time; the routing policy at time t; The set of flow information at time t is , Denote the number of flows, Denote the i-th flow information at time t; the flow information at time t is the set of queues and links through which the data sequence of the communication network at time t flows; the data sequence of the communication network at time t is determined according to the routing policy The i-th flow information at time t is denoted as , and respectively denote the starting queue and starting link through which the flow passes, and respectively denote the ending queue and ending link through which the flow passes;

[0077] The queue is used to store the data on the router port. In a communication network, a flow usually refers to a data sequence transmitted from a source node to one or more destination nodes in the network.

[0078] Based on the network information set, the network graph information is constructed, including:

[0079] S21, based on the network information set, construct a static influence graph set; the static influence graph set includes several static influence graphs; the expression of the static influence graph set is , is the number of time steps, Denote the static influence graph at the first time, Denote the static influence graph at the T-th time. Each static influence graph contains a node set and a feature set ; the node set includes node information and feature information; the node information includes flow information, link information, and queue information; the feature information includes flow node features, link node features, and queue node features; The corresponding flow node feature is , where is The size of the i-th flow at a moment, is the one-hot encoding of the priority type of the i-th flow, including EF, AFx, BE, etc., indicating that the i-th flow at time whether it can be correctly routed to the destination router, with a value of 1 indicating no and a value of 0 indicating yes; The corresponding link node feature is , where, represents the bandwidth of the link connecting router i and router j, is the one-hot encoding of the queue scheduling policy on the port connecting router i and router j, and the values of the one-hot encoding are first-in-first-out (FIFO) or differentiated services code point (DSCP);

[0080] The corresponding queue node feature is , . represents the queue size on the port connecting router i and router j, is the one-hot encoding of the queue priority type on the port connecting router i and router j, and its values are EF, AFx, BE, etc., represents the scheduling weight of the queue on the port connecting router i and router j.

[0081] S22, using the node type mapping function, map the node information in the static influence graph to node type information;

[0082] S23, using the relationship type mapping function, map the feature information in the network graph information to influence relationship information;

[0083] S24, using the node type information and influence relationship information, construct a dynamic influence graph;

[0084] S25, using the dynamic influence graph and the set of static influence graphs, construct the network graph information.

[0085] Both the static influence graph and the dynamic influence graph contain nodes and edges.

[0086] The node type mapping function The expression of is:

[0087] ,

[0088] where, represents the node information, is the node type information, and the value range of the node type information is , respectively represent the flow type, link type and queue type;

[0089] The relationship type mapping function has the following expression:

[0090] ,

[0091] where represents feature information, represents influence relationship information, and the value range of the influence relationship information is , is the link type flow type influence relationship information, is the queue type flow type influence relationship information, and so on for other influence relationship information.

[0092] The influence relationship information is a set of the corresponding two node features. Specifically, the directly connected links and queues form an influence relationship, which can be represented by a matrix; the influence relationship of the queues passing through on the link of the router is , and the influence relationship of the queues not passing through is , and influence relationships of and are formed between the flow nodes and the links and queues.

[0093] The network performance estimation model includes a heterogeneous graph multi-head attention layer, a temporal multi-head attention layer, and a multi-metric output layer;

[0094] The heterogeneous graph multi-head attention layer is used to process the network graph information to obtain an output vector; the input end of the heterogeneous graph multi-head attention layer is used to receive the network graph information; the output end of the heterogeneous graph multi-head attention layer is connected to the input end of the temporal multi-head attention layer;

[0095] The temporal multi-head attention layer is used to process the output vector of the heterogeneous graph multi-head attention layer to obtain a node embedding feature vector; the output end of the temporal multi-head attention layer is connected to the input end of the multi-metric output layer;

[0096] The multi-metric output layer is used to process the node embedding feature vector to obtain a predicted result value of the network performance; the output end of the multi-metric output layer is used to output the predicted result value of the network performance; the predicted result value of the network performance includes predicted result values of the flow throughput, flow delay, link utilization rate, and queuing delay of the communication network.

[0097] The heterogeneous graph multi-head attention layer includes a first processing model, a second processing model, and a third processing model;

[0098] The expression of the first processing model is:

[0099] ,

[0100] where, is the spatial transformation matrix corresponding to the type information, which is a parameter of the heterogeneous graph multi-head attention layer and can be changed during the training of the network performance estimation model, represents the output vector of the information of the i-th node in the static influence graph set after the k-th feature mapping, represents the feature of the i-th node of the j-th static influence graph in the static influence graph set, , represents the node type information obtained by inputting the information of the i-th node in the static influence graph set into the node type mapping function ;

[0101] The expression of the second processing model is:

[0102] ,

[0103] ,

[0104] where, represents the influence value of the j-th node on the i-th node through the edge of type r in the static influence graph, represents the vector splicing operation, is the attention parameter vector of the edge of type , which can be learned and changed during the training of the network performance estimation model, represents the position encoding value of the edge from the j-th node to the i-th node in the flow information, represents the -th node corresponding to the weight coefficient of the j-th node, which can be learned and changed during the training of the network performance estimation model, represents the node connected to the neighbor nodes through the edge of type , represents the activation function;

[0105] The expression of the third processing model is:

[0106] ,

[0107] where, , is the type value of the node , , ​is the set of types of relevant edges of nodes with node type . represents the edge type variable, and is the value range of . For example, if the node type is , then represents the output vector after the (k + 1)-th feature mapping of the information of the i-th node in the static influence graph set, represents the output vector after the k-th feature mapping of the information of the j-th node in the static influence graph set; head represents the attention head number, is the number of attention heads of the heterogeneous graph multi-head attention layer, is an injective function, implemented using a single-layer multi-layer perceptron, and is used to map the information of the influence relationships of different types of nodes and the multi-head attention information to the feature space of the same dimension; The dimensionality transformation process of is

[0108] The relevant edges of a node refer to the influence relationship information corresponding to the feature information related to the node information. For example, the relevant feature information of node F is , .

[0109] The input sequence of the temporal multi-head attention layer is , , where is the output result of the heterogeneous graph multi-head attention layer at time t for the type value . The length of the input sequence is the time window , is the number of nodes with node type value , is the dimension of the node features.

[0110] The -th moment output of the temporal multi-head attention layer is , is the new node feature embedding after fusing the historical time feature , with a dimension of , which is a preset value.

[0111] The temporal multi-head attention layer includes a fourth processing model and a fifth processing model; the processing process of the fourth processing model includes:

[0112] Perform an update operation on the input sequence, and the expression of the update operation is:

[0113] ,

[0114] where is an encoding function for the value of the current moment t, is a sin encoding function; for each in perform an update operation to obtain an updated input sequence;

[0115] The processing process of the fifth processing model includes:

[0116] ,

[0117] ,

[0118] ,

[0119] ,

[0120] where , is an attention weight matrix determined by the scaled dot - product attention function, is the attention weight between the moment and the p moment, is the time influence degree value between the t moment and the k moment, where k is the moment value of the time window, , is the time influence degree value between the t moment and the p moment, is the vector dimension of the node embedding feature vector; is a MatMul network parameter, and the query matrix of the type value is , and this parameter can be learned and changed during the training of the network performance estimation model. The key matrix is , and this parameter can be learned and changed during the training of the network performance estimation model. The value matrix is , and this parameter can be learned and changed during the training of the network performance estimation model, is the number of time attention heads of the temporal multi - head attention layer, is a preset mapping matrix, and this parameter can be learned and changed during the training of the network performance estimation model, represents the node embedding feature vector obtained after fusing historical time features, and Concat represents the vector concatenation operation, represents the output result of the first temporal attention layer at the t moment for the type value , represents the element of the t - th row and p - th column of the matrix.

[0121] The fifth processing model includes several temporal attention layers.

[0122] The multi-index output layer includes a first multi-layer perceptron module, a second multi-layer perceptron module, a third multi-layer perceptron module, and a fourth multi-layer perceptron module;

[0123] The calculation expressions of the first multi-layer perceptron module, the second multi-layer perceptron module, the third multi-layer perceptron module, and the fourth multi-layer perceptron module are respectively:

[0124] ,

[0125] ,

[0126] ,

[0127] ,

[0128] where represents the predicted result values of the flow throughput, flow delay, link utilization rate, and queuing delay of the communication network at time t, , , , respectively represent the multi-layer perceptron operations of the first multi-layer perceptron module, the second multi-layer perceptron module, the third multi-layer perceptron module, and the fourth multi-layer perceptron module;

[0129] The information of the i-th node in the static influence graph set can be the information of the i-th node in the j-th static influence graph in the static influence graph set;

[0130] The sin coding function can be implemented by the sin function;

[0131] The scaled dot-product attention function can be implemented by the scaled dot-product attention module.

[0132] The position encoding value , is calculated through calculation.

[0133] is the quotient of the position subscript of the vector divided by 2 and taking the floor, , for example: is 32-dimensional, The value of

[0134] Table 1 Value schematic table

[0135]

[0136] The network performance estimation model can be implemented based on Python 3.7.10 and above versions, trained on the Huawei Ascend Atlas 300 graphics card. The training samples can be obtained by annotating the communication network structure information and the measured communication network performance indicators. When training, the sample batch size is not less than 10, the number of iteration rounds is not less than 400 times, the optimizer is Adam, and the learning rate is not greater than 1×10 -4 , and the loss function can be the cross-entropy loss function. The trained model can be evaluated using accuracy, precision, and recall, which are not limited in the embodiments of the present invention.

[0137] In the second aspect of the embodiments of the present invention, a data-driven communication network performance estimation method is disclosed, which specifically includes the following steps:

[0138] S1: Capture the mutual influence relationships among various attributes in the communication network and construct an influence graph model; specifically,

[0139] S11: Characterize the communication network and construct a communication network model; specifically, this method defines the communication network at time as ,

[0140] where the router set ,

[0141] the link set ,

[0142] the flow set at time , the queue set , the router routing policy set , is the routing policy of router at time. This method defines a flow as a tuple sequence of queues and links determined according to the routing policy and the start and end queue links , representing the queues and links through which the flow passes.

[0143] S12: Abstract the key elements affecting the communication network performance into nodes and extract their key features; specifically,

[0144] This method defines the temporal influence graph to be generated as a series of static influence graph snapshots, , where is the number of time steps, and each snapshot is a static heterogeneous graph, defined as , including the node set and the influence relationship set . Nodes include multiple types, and the node type mapping function is defined as , node type includes flows, links, and queues. The influence relationships include multiple types, and the influence relationship type mapping function is defined as , since the influence relationships are different among different node types. For example, link bandwidth has a greater impact on the performance of flows, the size and burstiness of flows affect the queuing delay, and the queue scheduling strategy affects the link utilization, etc. Therefore, this method defines six different types of influence relationships among different nodes , is the influence of the link on the flow . The other influence relationship types can be deduced by analogy.

[0145] For different types of nodes, this method extracts the key features that have the greatest impact on network performance for modeling. It should be noted that when users are concerned about other features, they can modify and expand it according to needs.

[0146] The features of the flow node are

[0147] . Among them, is the size of the flow at time is the one-hot encoding of the flow priority type, including EF, AFx, BE, etc. indicates whether the flow can be correctly routed to its destination at time , 1 indicates error, and 0 indicates correct delivery.

[0148] The features of the link node are . Among them represents the link bandwidth, while is the one-hot encoding of the queue scheduling strategy on the port connecting node and node , including first-in-first-out (FIFO) or differentiated services code point (DSCP).

[0149] The features of the queue node are . is modeled as a feature vector . represents the queue size, is the one-hot encoding of the queue priority type (such as EF, AFx, BE, etc.), while represents the queue scheduling weight.

[0150] S13: Construct the influence relationships among flow, link, and queue elements to form an influence graph; specifically, the influence relationship refers to the influence relationship among flow, link, and queue nodes, including , and Three types of two-way influence relationships. First, obtain the source node and destination node of the flow, and then based on the topological structure and routing policy obtain the links and queues that the flow passes through. The directly connected links and queues form an influence relationship. An influence relationship is formed between the flow nodes and the links and queues and an influence relationship. If the flow is not connected to the destination node, the flow is unreachable, denoted as ; otherwise, . It should be noted that this method regards the network topological structure as implicit information, and only constructs the direct link, queue relationships related to the flow and the relationships between these links and queues. This eliminates the nodes irrelevant to the flow performance, thus reducing the scale of the influence graph and improving the training speed and prediction accuracy.

[0151] S2: To learn the influence relationships between different types of nodes in the heterogeneous influence graph, a heterogeneous graph multi-head attention mechanism (HGMA) is designed. It first uses the edge attention mechanism to learn the influence of different types of nodes. Then, the influence results of different types of nodes are aggregated to form the final node embedding. Then, sub-sampling is performed to extract the indirect influence between elements; specifically,[[]]

[0152] S21: Map the features of different types of nodes to the same feature space; specifically, first, for any influence graph this method extracts its node features as the initial embedding . represents the type of node , is the feature dimension of the node of type . Due to the heterogeneity of nodes, different types of nodes have different feature spaces. This method uses a specific transformation matrix

[0153]

[0154] to map the features of different types of nodes to the same feature space. The calculation expression of S21 is: S22: Use the attention mechanism to learn the weights between different edge types; specifically, given a pair of nodes connected by an edge type , the edge attention is , which represents the influence degree of node on node through the edge type

[0155]

[0156] where Represents a connection operation, , which is the edge type of the edge sharing attention vector. It should be noted that due to the different performance characteristics of the link and the queue at different positions in the flow. Therefore, it is necessary to encode the relative positions of the link and the queue with respect to the flow When the node and the node types are not both When . In this work, the commonly used sine and cosine functions with different frequencies are used to encode the positions as follows

[0157] ,

[0158] ,

[0159] where is the position of the link or queue in the flow, is the dimension. Then, after obtaining the influence degree between node pairs based on the edge type, this method normalizes them through the softmax function to obtain the weight coefficients ,

[0160] ,

[0161] where represents the neighbor of node based on the edge type , represents the activation function. Therefore, 's weight coefficients depend on their features. Then, the edge-type based embedding of node can be aggregated through the projected features of its neighbors.

[0162] S23: Design a multi-head attention mechanism to obtain spatial features from multiple perspectives; specifically, at the same time, due to the scale-free property of the heterogeneous influence graph, the high variance of the graph data leads to unstable training. Therefore, this method introduces a multi-head attention mechanism, taking the learned features as the embedding of the fused features, and each attention head has independent parameter weights. The expression of step S23 is:

[0163] ,

[0164] where is the type of node , , is the relevant edge type of node type , for example, if the node type is , then . is the number of attention heads in the heterogeneous graph multi-head attention layer, is an injective function implemented using a single layer of MLP, which maps the influence results of different types of nodes and the multi-head attention results to the same feature space .

[0165] S3: To further capture the temporal evolution characteristics of the dynamic influence graph, this method designs a temporal multi-head attention layer, which realizes the node feature embedding at the current moment by fusing the historical time series features; specifically,

[0166] S31: Aggregate the temporal position features and the initial node features; specifically, since there are multiple types in the network, this method uses different attention parameters to learn the temporal evolution features of different types of nodes. The input to this layer is the historical time series features of a specific node type at a moment . This feature has fully captured the spatial information of the influence graph, separating the structure modeling and the temporal modeling. Specifically, for the node type , this method defines the input as , where is the final embedding vector after capturing the spatial information of the influence graph, the sequence length is the time window , the number of nodes of node type , is the node feature dimension. The output is at moment the new node feature embedding after fusing the historical time features , with the dimension of . The update expression of the embedding vector is:

[0167] .

[0168] The key objective of the temporal multi-head self-attention layer is to capture the influence of the features within the historical time window on the network embedding at the current moment. is the representation after fusing the spatial features at moment . This method uses as the query to focus on its historical representation. Different from the spatial attention that acts on the adjacent node representations, the temporal attention fully depends on the time window features of each node, thus promoting the effective parallelism between nodes.

[0169] S32: Design a multi-head self-attention strategy to fuse historical time-series features to achieve the embedding of time features. Specifically, this method designs a Scaled Dot-Product Attention and Multi-Head Attention strategy, where the query, key, and value are set to the node input representation. The query vector is , the key is and the value is . The definition of the time self-attention function is:

[0170] ,

[0171] ,

[0172] ,

[0173] where is the attention weight matrix obtained by the multiplicative attention function, is the moment vector and the attention weight at the moment. To capture different types of time dependencies between different features, this method uses a multi-head attention mechanism to learn the time dependencies of the input sequence from multiple perspectives. Each "head" can learn different patterns and dependencies, thus providing richer information. The expression of this process is:

[0174] ,

[0175] where is the number of time attention heads, is the mapping matrix that maps it to a unified feature space.

[0176] S32: Finally, use different readout functions to achieve multi-granularity network performance estimation. The readout function is a multi-layer perceptron.

[0177] A communication network is a collection of a group of devices and protocols used to transmit information. These devices include, but are not limited to, routers, switches, servers, terminal devices (such as mobile phones, computers), etc., and the protocols are the standards that stipulate how these devices transmit data. Communication networks can be classified according to different criteria, such as coverage range, transmission medium, network topology, etc.

[0178] Those skilled in the art can understand that to implement all or part of the processes of the above embodiments, it can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a disk, an optical disc, a read-only memory, or a random access memory, etc.

[0179] The above are only embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, various modifications and variations can be made to the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the scope of the claims of the present invention.

Claims

1. A data-driven communication network performance prediction method, characterized in that: include: S1, obtain network information set; S2, constructing network graph information based on the network information set; S3, using the trained network performance estimation model to process the network graph information to obtain a predicted result value of the network performance; The step of constructing network diagram information based on the network information set includes: S21, constructing a static influence diagram set based on the network information set; the static influence diagram set includes a plurality of static influence diagrams; The expression of the static influence diagram set is , is the number of time steps, represents the static influence diagram at the first moment, Represents the static influence diagram at the Tth moment. Each static influence diagram contains a set of nodes and feature set ; The node set includes node information and feature information; the node information includes flow information, link information and queue information; the feature information includes flow node features, link node features and queue node features; S22, mapping the node information in the static influence diagram into node type information using a node type mapping function; S23, mapping the characteristic information in the static influence graph into influence relationship information using a relationship type mapping function; S24, constructing a dynamic influence diagram using the node type information and the influence relationship information; S25, constructing network diagram information using the dynamic influence diagram and the static influence diagram set; The network performance estimation model includes a heterogeneous graph multi-head attention layer, a temporal multi-head attention layer and a multi-indicator output layer; The heterogeneous graph multi-head attention layer is used to process the network graph information to obtain an output vector; the input end of the heterogeneous graph multi-head attention layer is used to receive the network graph information; the output end of the heterogeneous graph multi-head attention layer is connected to the input end of the temporal multi-head attention layer; The temporal multi-head attention layer is used to process the output vector of the heterogeneous graph multi-head attention layer to obtain a node embedding feature vector; the output end of the temporal multi-head attention layer is connected to the input end of the multi-index output layer; The multi-indicator output layer is used to process the node embedded feature vector to obtain the predicted result value of the network performance; the output end of the multi-indicator output layer is used to output the predicted result value of the network performance; the predicted result value of the network performance includes the predicted result values ​​of the flow throughput, flow delay, link utilization and queuing delay of the communication network.

2. The data-driven communication network performance prediction method according to claim 1, characterized in that: The network information set includes communication network information at each moment; The communication network information at a certain moment is defined as , where the router information set is represented as , is the number of routers, represents the i-th router, and the link information set is expressed as , represents the i-th link information, Represents the total number of links, and the queue information set is represented as , Indicates the i-th queue information, is the total number of queues, and the routing strategy set of the router at time t is expressed as , It is a router exist Routing strategy at the moment; The flow information set at time is , Indicates the number of flows, represents the i-th flow information at time t; the flow information at time t is the set of queues and links through which the data sequence of the communication network at time t flows; the data sequence of the communication network at time t, according to the routing strategy Determined; the i-th flow information at time t is expressed as , and They represent the starting queue and starting link through which the flow flows. and They respectively represent the termination queue and termination link through which the flow flows.

3. The data-driven communication network performance prediction method according to claim 1, characterized in that: The node type mapping function The expression is: ， in, Indicates node information. is the node type information, the value range of the node type information is , They represent flow type, link type and queue type respectively; The relationship type mapping function The expression is: ， in, Represents feature information, Indicates the influence relationship information. The value range of the influence relationship information is , Link Type Convection Type The impact relationship information, For queue type Convection Type The impact relationship information.

4. The data-driven communication network performance prediction method according to claim 1, characterized in that: The heterogeneous graph multi-head attention layer includes a first processing model, a second processing model and a third processing model; the second processing model is connected to the first processing model and the third processing model respectively; The calculation expression of the first processing model is: ， in, Node type information The corresponding space transformation matrix is, Represents the output vector of the i-th node information of the static influence diagram set after the k-th feature mapping, represents the characteristics of the i-th node of the j-th static influence diagram in the static influence diagram set, , Indicates that the information of the ith node in the static influence diagram set is input into the node type mapping function The node type information obtained later; The expression of the second processing model is: ， ， in, Represents the influence value of the j-th node on the i-th node through the edge of type r in the static influence graph, represents the vector concatenation operation, Represents the output vector of the j-th node information of the static influence diagram set after the k-th feature mapping, Is of type The attention parameter vector of the edge, Represents the position encoding value of the edge from the jth node to the ith node in the flow information, Indicates The weight coefficient of the jth node corresponds to the node. Representation Node The pass type is The neighbor nodes connected by the edge, represents the activation function; The expression of the third processing model is: ， in, , For Node The type value of , The node type is The set of types of edges related to the nodes, represents the edge type variable, for The value range of It represents the output vector of the i-th node information of the static influence diagram set after the k+1-th feature mapping. It represents the output vector of the jth node information of the static influence graph set after the kth feature mapping; head represents the sequence number of the attention head, is the number of attention heads in the heterogeneous graph multi-head attention layer, It is an injective function, implemented using a multi-layer perceptron, which is used to map different types of node influence relationship information and multi-head attention information into a feature space of the same dimension.

5. The data-driven communication network performance prediction method according to claim 4, characterized in that: The input sequence of the temporal multi-head attention layer is ,in is the type value of the heterogeneous graph multi-head attention layer at time t The output result of the input sequence length is the time window ; The temporal multi-head attention layer includes a fourth processing model and a fifth processing model; The fourth processing model is connected to the fifth processing model; The processing process of the fourth processing model includes: Perform an update operation on the input sequence, the expression of the update operation is: ， in, is the encoding function for the value at the current time t, is the sin encoding function; right Each of Both perform update operations to obtain updated input sequences; The processing process of the fifth processing model includes: ， ， ， ， in, is the attention weight matrix determined by the scaled dot-product attention function, for The attention weights at time t and time p, is the time influence value between time t and time k, k is the time value of the time window, , is the time influence value between time t and time p, The vector dimension of the node embedding feature vector, type value The query matrix is , type value The bond matrix is , type value The value matrix of , is the number of temporal attention heads of the temporal multi-head attention layer, is the preset mapping matrix, represents the node embedding feature vector obtained after integrating historical time features, Concat represents the vector concatenation operation, Indicates the type value at time t The output of the first temporal attention layer, and so on. Indicates the type value at time t No. The output of the temporal attention layer is Representation Matrix The element at the tth row and pth column of .

6. A data-driven communication network performance prediction device, characterized in that: The device comprises: A memory storing executable program code; a processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the data-driven communication network performance prediction method according to any one of claims 1 to 5.

7. A computer storable medium, characterized in that: The computer storable medium stores computer instructions, and when the computer instructions are called by a computer, they are used to execute the data-driven communication network performance prediction method according to any one of claims 1 to 5.

8. An information data processing terminal, characterized in that: The information data processing terminal is used to implement the data-driven communication network performance prediction method as described in any one of claims 1 to 5.

Images