Monitoring data recovery model training method, recovery method, device and storage medium

Abstract

The application discloses a kind of monitoring data recovery model training method, recovery method, equipment and storage medium, the method includes obtaining N rounds of historical measurement event data, selects continuous K rounds of historical measurement event data to construct historical measurement event data set, obtains N-K+1 historical measurement event data set;K-1 rounds of historical measurement event data are constructed bipartite graph to the front, to obtain the bipartite graph sequence of each historical measurement event data set, the Kth round of historical measurement event data is converted into target matrix;Streaming network monitoring data recovery model is constructed;Bipartite graph sequence is used as the input of the context information extraction module of model, the Kth round of historical measurement time data is used as the input of the event coding module of model, corresponding target matrix is used as model output, streaming network monitoring data recovery model is trained.The present application can realize online data recovery in real time, event by event.

Description

Technical Field

[0001] This invention belongs to the field of data recovery technology, and particularly relates to a training method for a real-time event-based streaming network monitoring data recovery model (ResNMMF), a streaming network monitoring data recovery method, device, and storage medium. Background Technology

[0002] Network monitoring is crucial for network operation and management, essential for tasks such as anomaly detection, network routing planning, congestion control, and network troubleshooting. However, monitoring the entire network is challenging for the following reasons:

[0003] First, for a network with N nodes, the time complexity of measuring all end-to-end source-destination pairs (OD pairs) in one round of measurement is O(N²); when a network requires multiple rounds of measurement, the cumulative complexity becomes unacceptable.

[0004] Secondly, network measurement inevitably incurs additional overhead, which can affect network system performance and lead to observer effects. Reducing the number of packets used for network monitoring can mitigate unreliable measurement results caused by observer effects. Therefore, a sparse sampling mechanism that measures only a subset of all OD pairs is commonly employed. To record sparsely sampled network monitoring data for the entire network in a single measurement round, a sparse network monitoring matrix (NMM) is formed, with its rows and columns representing the origin and destination, respectively. With advancements in sparse reconstruction techniques, NMM recovery using low-rank matrix completion has been proposed; however, matrix completion-based methods focus on a single NMM and do not consider temporal dynamics, thus potentially compromising recovery accuracy. Recent research has used a series of NMMs arranged chronologically to form a high-order array (also known as a tensor) and applied tensor completion methods to recover missing terms, extracting more complex correlations, including temporal correlations. Therefore, tensor-based methods generally achieve better accuracy than matrix-based methods.

[0005] Some researchers have proposed using the low-rank properties of network data to recover missing entries. Matrix completion-based algorithms cannot capture temporal dynamics, resulting in insufficient data recovery accuracy. Introducing tensor completion methods can better utilize the implicit spatiotemporal correlations in sparsely collected network monitoring data. LTC (see Xie K, Wang X, Wang X, et al. Accurate recovery of missing network measurement data with localized tensor completion[J].IEEE / ACM Transactions on Networking,2019,27(6):2222-2235.) decomposes the tensor using locality-sensitive hashing to construct multiple low-rank sub-tensors with strong internal correlations, thereby improving accuracy. Reshape-Align (see Xie K, Peng C, Wang X, et al. Accurate recovery of internet trafficdata under variable rate measurements[J]. IEEE / ACM transactions on networking, 2018, 26(3):1137-1150.) proposes reshaping measurement matrices with variable rate measurements and combining them into time-aligned regular tensors to achieve more practical and accurate missing data imputation. A major drawback of tensor completion is its high computational complexity despite its high accuracy. Some works focus on accelerating matrix / tensor completion using GPUs (Graphics Processing Units). However, it is still impossible to recover missing entries in milliseconds. Furthermore, GPU-accelerated methods require hardware support and are not energy efficient.

[0006] With the advancement of deep learning, some researchers have proposed data completion algorithms based on neural networks. DMF (see Xue HJ, Dai X, Zhang J, et al. Deep matrix factorization models for recommendersystems[C] / / IJCAI.2017,17:3203-3209.) proposes using neural networks to improve user and item embeddings for accurate recommendations; DLMC (see Fan J, Chow T. Deep learning based matrix completion[J].Neurocomputing,2017,266:540-549.) applies an encoder-decoder architecture to recover missing entries; GAIN (see Yoon J, Jordon J, Schaar M. Gain: Missing data imputation using generative adversarial nets[C] / / International conference on machine learning.PMLR,2018:5689-5698.) demonstrates that generative adversarial networks are also a potential and powerful tool for sparse matrix imputation. To fully leverage the nonlinear expressive power of neural networks to explore nonlinear data correlations, CoSTCo (see Liu H, Li Y, Tsang M, et al. Costco: A neural tensor completion model for sparse tensors [C] / / Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 2019: 324-334.) and NTC (see Xie K, Lu H, Wang X, et al. Neural tensor completion for accurate network monitoring [C] / / IEEE INFOCOM 2020-IEEE Conference on Computer Communications. IEEE, 2020: 1688-1697.) utilize convolutional neural networks (CNNs) to design better interaction functions to fill sparse tensors. However, neural networks are typically trained on a fixed-size historical dataset. Without efficient and effective update algorithms, even considering online learning strategies, real-time data recovery tasks cannot be supported.

[0007] In fact, network measurement events can be viewed as tuples describing the start point, end point, start time, and result of the measurement. Current schemes typically convert these events into sparse matrices or assume that the observed entries in the sparse matrix arrive simultaneously, then apply matrix completion to recover the missing data. Tensor completion algorithms handle this similarly. Because these methods collect events at time intervals in a batch, ignoring the low-latency nature of event data, matrix / tensor completion methods are called batch-based network monitoring. Applying matrix or tensor completion to streaming event-based network monitoring systems suffers from computational waste and high latency. As events continuously arrive, entries in the NMM are modified, and even if only a few entries change, the matrix / tensor completion model requires a significant amount of time to retrain, rather than simply calculating these changes. Therefore, current matrix-based modeling is unsuitable for real-time data recovery tasks and time-sensitive network applications such as network anomaly detection and automated network troubleshooting, as these tasks require complete NMM data for real-time decision-making.

[0008] Currently, no other algorithm can achieve online data recovery using a real-time, event-by-event processing approach.

[0009] Definition: Streaming refers to the measurement of events in the form of a continuous data stream. Summary of the Invention

[0010] The purpose of this invention is to provide a monitoring data recovery model training method, recovery method, device, and storage medium to solve the problem that traditional algorithms cannot achieve online recovery of network monitoring data in a real-time, event-by-event processing manner.

[0011] This invention solves the above-mentioned technical problems through the following technical solution: a training method for a streaming network monitoring data recovery model, the method comprising:

[0012] Acquire N rounds of historical measurement event data, where each round of historical measurement event data includes n historical measurement events, and each historical measurement event includes source node ID, target node ID, timestamp, and measurement result;

[0013] A historical measurement event dataset is constructed by selecting K consecutive historical measurement event data from N rounds of historical measurement event data, thus obtaining N-K+1 historical measurement event datasets;

[0014] A bipartite graph is constructed for the first K-1 rounds of historical measurement event data in each historical measurement event dataset, thereby obtaining a bipartite graph sequence for each historical measurement event dataset; the Kth round of historical measurement event data in each historical measurement event dataset is transformed into a target matrix;

[0015] A streaming network monitoring data recovery model is constructed, comprising a context information extraction module, an event encoding module, and a data recovery module; the context information extraction module and the event encoding module are respectively connected to the data recovery module.

[0016] The bipartite graph sequence of each historical measurement event dataset is used as the input to the context information extraction module, the Kth round of historical measurement time data of the corresponding historical measurement event dataset is used as the input to the event encoding module, and the corresponding target matrix is ​​used as the model output. The streaming network monitoring data recovery model is trained to obtain the target streaming network monitoring data recovery model.

[0017] Furthermore, a bipartite graph is constructed for the first K-1 rounds of historical measurement event data in each of the historical measurement event datasets, specifically including:

[0018] For the k-th round of historical measurement event data, an undirected edge is created between the source node and the target node of each historical measurement event. The value of the undirected edge is the measurement result of the historical measurement event, thus obtaining a bipartite graph of the k-th round of historical measurement event data; where k = 1, 2, ..., K-1.

[0019] Furthermore, the specific implementation process for training the streaming network monitoring data recovery model includes:

[0020] The context information extraction module is used to extract features from the bipartite graph sequence to obtain the source node context feature vector and the target node context feature vector.

[0021] The event encoding module is used to extract features from the corresponding Kth round of historical measurement event data to obtain the source node state vector and the target node state vector;

[0022] The data recovery module is used to perform deep feature extraction on the source node context feature vector and source node state vector, the target node context feature vector and target node state vector to obtain the state prediction matrix of the Kth round of historical measurement event data;

[0023] A loss function is constructed based on the state prediction matrix and the target matrix. The parameters of the streaming network monitoring data recovery model are adjusted with the goal of minimizing the loss function, thereby achieving model training.

[0024] Furthermore, the context information extraction module is used to extract features from the bipartite graph sequence to obtain the source node context feature vector and the target node context feature vector, specifically including:

[0025] For each bipartite graph in the bipartite graph sequence, an initial embedding matrix of source nodes is randomly generated based on the number of source nodes corresponding to the bipartite graph. An initial embedding matrix of target nodes is randomly generated based on the number of target nodes in the bipartite graph. Where |S| is the number of source nodes, |D| is the number of target nodes, dim is the embedding dimension, and R is the real number space;

[0026] Based on the initial embedding matrix of the source node and the initial embedding matrix of the target node Feature extraction is performed on the bipartite graph to obtain the source node factor matrix and the target node factor matrix;

[0027] Feature extraction is performed on the factor matrices of all source nodes in the bipartite graph sequence to obtain the source node context feature vector A. ctx Feature extraction is performed on the factor matrices of all target nodes in the bipartite graph sequence to obtain the target node context feature vector B. ctx The specific formula is as follows:

[0028] A ctx =GRU({S 1 ,S 2 ,…,S (K-2) ,S (K-1)})

[0029] B ctx =GRU({D 1 D 2 ,…,D (K-2) D (K-1)})

[0030] Among them, S 1 ,S 2 ,…,S (K-2) ,S (K-1) G is a bipartite graph. 1 G 2 ,…,G (K-2) G (K-1) The source node factor matrix, D 1 D 2 ,…,D (K-2) D (K-1) G is a bipartite graph. 1 G 2 ,…,G (K-2) G (K-1) The target node factor matrix, GRU represents a recurrent neural network.

[0031] Furthermore, based on the initial embedding matrix of the source node and the initial embedding matrix of the target node Feature extraction is performed on the bipartite graph to obtain the source node factor matrix and the target node factor matrix, specifically including:

[0032] For any edge (s, d) in a bipartite graph, take the initial embedding matrix of the source node. The s-th row is used as the initial embedding feature s (0) Take the initial embedding matrix of the target node. The d-th row is used as the initial embedding feature d (0) Where s is the starting node ID of edge (s,d), and d is the ending node ID of edge (s,d).

[0033] The initial embedding features s of the edge (s,d) of the bipartite graph (0) Feature encoding and feature aggregation are performed to obtain the embedding features of the starting node of edge (s,d); the initial embedding features d of edge (s,d) of the bipartite graph are obtained. (0) Feature encoding and feature aggregation are performed to obtain the embedding features of the endpoint node of edge (s,d); subsequently, the embedding features of the starting node and the endpoint node of each edge in the bipartite graph are obtained; the specific formula is as follows:

[0034]

[0035]

[0036] Where, r (L) Let r be the embedding feature of node r in layer L. (L-1) Let g be the embedding feature of node r in the (L-1)th layer, and g() be a trainable linear transformation function. Let r be the set of all nodes that point to node r. Let W be the intermediate feature encoding vector, W be the trainable weight matrix, b be the trainable bias matrix, and u be the intermediate feature encoding vector. (L-1) e represents the embedding feature of node u in the (L-1)th layer. ur Let (u,r) be the value of the edge.

[0037] The initial embedding matrix of the source nodes is replaced with the embedding features of all starting nodes in the bipartite graph at layer L. The corresponding rows are used to obtain the source node factor matrix. The initial embedding matrix of the target node is then replaced with the embedding features of all endpoint nodes in the bipartite graph at layer L. The corresponding rows yield the target node factor matrix.

[0038] Furthermore, the event encoding module is used to extract features from the corresponding Kth round of historical measurement event data to obtain the source node state vector and the target node state vector. The specific implementation process includes:

[0039] An initial learning matrix for source nodes is randomly generated based on the number of source nodes in the historical measurement event data of the Kth round. An initial learning matrix for target nodes is randomly generated based on the number of target nodes in the historical measurement event data of the Kth round. Where |S| is the number of source nodes, |D| is the number of target nodes, dim is the embedding dimension, and R is the real number space;

[0040] Define the source node memory matrix M S and the target node memory matrix M D Source node memory matrix M S The number of rows and columns and the initial learning matrix of the source node The same applies to the target node memory matrix M. D The number of rows and columns and the initial learning matrix of the target node The same, and the source node memory matrix M S and the target node memory matrix M D Initialize to 0;

[0041] For the measurement event e = (p, q, t, v) in the historical measurement event data of the Kth round, take the initial learning matrix of the source node. The p-th row is taken as P, and the initial learning matrix of the target node is taken. The q-th row is taken as Q; where p is the source node ID, q is the target node ID, t is the timestamp, and v is the measurement result;

[0042] Memory matrix M of source nodes S and the target node memory matrix M D To update, the specific update formula is as follows:

[0043]

[0044]

[0045] Among them, Updater k (·) represents a single-layer network, Updater k (x)=ReLU(W k x+b k ), W k and b k For learnable parameters, ReLU(x) = max(x, 0); h = [P; Q; v]; M is the source node memory matrix after the i-th update. S , Let M be the memory matrix of the target node after the i-th update. D i = 1, 2, ..., M, where the update count M is the number of measurement events in the historical measurement event data of the Kth round;

[0046] The updated source node memory matrix is ​​used with a max pooling layer. and target node memory matrix After processing, the source node state vector A is obtained. curr and the target node state vector B curr The specific formula is as follows:

[0047]

[0048]

[0049] MaxPooling() is the maximum pooling layer.

[0050] Furthermore, the specific expression for the loss function is as follows:

[0051]

[0052] in, For loss function, Let X be the state prediction matrix. Ω Let |Ω| be the target matrix, and |Ω| be the target matrix X. Ω The number of non-zero elements in the middle, ||·|| F Let λ denote the Frobenius norm, where λ is a constant.

[0053] Based on the same concept, the present invention also provides a method for recovering streaming network monitoring data, comprising the following steps:

[0054] Acquire network measurement event data from the previous K-1 rounds and network measurement event data from the current round;

[0055] Construct a bipartite graph for each round of network measurement event data in the first K-1 rounds of network measurement event data;

[0056] The target streaming network monitoring data recovery model is invoked, and the target streaming network monitoring data recovery model is trained according to the streaming network monitoring data recovery model training method described above;

[0057] Based on the target streaming network monitoring data recovery model, the current round of network measurement event data is recovered to obtain the state prediction matrix of the current round of network measurement event data. The state prediction matrix is ​​the recovered network monitoring data.

[0058] Based on the same concept, the present invention provides an electronic device, comprising:

[0059] Memory, used to store computer programs;

[0060] A processor is used to implement the streaming network monitoring data recovery model training method or streaming network monitoring data recovery method as described above when executing the computer program.

[0061] Based on the same concept, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the streaming network monitoring data recovery model training method or the streaming network monitoring data recovery method as described above.

[0062] Beneficial effects

[0063] Compared with the prior art, the advantages of the present invention are as follows:

[0064] This invention proposes an asynchronous processing approach combining online processing (context information extraction module) and offline processing (event encoding module). The context information extraction module captures the spatiotemporal correlation information of sparse data, while the event encoding module processes measurement event data that is not sensitive to order. This separation of online and offline processing improves the model response speed when deployed in measurement scenarios. Compared to traditional online learning techniques, this invention eliminates the need for online gradient descent, significantly reducing processing latency and enhancing parallelism. Unlike traditional tensor filling and matrix filling, which only focus on the characteristics of existing data, this invention does not require retraining the model to process new data. The features learned from historical data can be directly adapted to unlearned features. This invention enables real-time, event-by-event online recovery of network monitoring data. Attached Figure Description

[0065] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0066] Figure 1 This is a flowchart of the streaming network monitoring data recovery model training method in an embodiment of the present invention;

[0067] Figure 2 This is a comparison of the ResNMMF model and the mathematical low-rank tensor filling algorithm on different datasets in the embodiments of this invention;

[0068] Figure 3 This is a comparison of the ResNMMF model and the neural network tensor filling algorithm on different datasets in the embodiments of this invention;

[0069] Figure 4 This is a comparison of the ResNMMF model and STF on different datasets in the embodiments of this invention. Detailed Implementation

[0070] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] The technical solutions of this application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0072] Example 1

[0073] like Figure 1 As shown in the figure, the streaming network monitoring data recovery model training method provided by the embodiment of the present invention includes the following steps:

[0074] Step 1: Obtain historical measurement event data for N rounds.

[0075] Each round of historical measurement event data includes n historical measurement events. Each historical measurement event includes source node ID, target node ID, timestamp, and measurement result. Each historical measurement event is represented by a quadruple, denoted as e = (p, q, t, v), where p is the source node ID, q is the target node ID, t is the timestamp, and v is the measurement result.

[0076] Assuming that network measurement or monitoring is performed in one round every minute, then N rounds of historical measurement event data are the measurement event data of the past N minutes, and the measurement event data of each minute includes n measurement events.

[0077] Step 2: Construct a historical measurement event dataset.

[0078] A historical measurement event dataset is constructed by selecting K consecutive historical measurement event data from N rounds of historical measurement event data, thus obtaining N-K+1 historical measurement event datasets, where K < N.

[0079] To capture network state information that changes over time, the dataset is divided into multiple historical measurement event datasets based on the timestamp t of each measurement event. Each historical measurement event dataset includes K rounds of historical measurement event data, denoted by E, with its superscript indicating the measurement round. A single historical measurement event dataset is then denoted as:

[0080] {E 1 E 2 ,…,E (K-2) E (K-1) E (K)}

[0081] For a historical measurement event dataset, E 1 E 2 ,…,E (K-2) E (K-1) E is the historical measurement event data from the previous K-1 rounds. (K) This represents the historical measurement event data for the Kth round.

[0082] Step 3: Data Processing

[0083] In the specific application of the ResNMMF model, in order to improve the online processing capability of the ResNMMF model, the input is divided into two parts: one part is used for offline processing and the other part is used for online processing. Therefore, during training, the data in each historical measurement event dataset is also divided into two parts: the first K-1 rounds of historical measurement event data are used for offline processing, and the Kth round of historical measurement event data is used for online processing.

[0084] For each historical measurement event dataset, the historical measurement event data of each round in the first K-1 rounds (e.g., the historical measurement event data of the kth round, k = 1, 2, ..., K-1) is transformed into a bipartite graph. The specific implementation process is as follows:

[0085] Create an undirected edge between the source node p and the target node q for each historical measurement event, with the value e of the undirected edge. pq Let v be the measurement result of this historical measurement event, from which we obtain the bipartite graph G of the historical measurement event data in the kth round. (k) This yields a bipartite graph sequence for each historical measurement event dataset, denoted as {G}. 1 G 2 ,…,G (K-2) G (K-1)}

[0086] For each historical measurement event dataset, the historical measurement event data of the Kth round is transformed into a target matrix. The specific implementation process is as follows:

[0087] The Kth round of historical measurement event data includes n historical measurement events. Let v be the value in the p-th row and q-th column of the matrix. This gives us the target matrix, where p is the source node ID and q is the target node ID.

[0088] Step 4: Construct a streaming network monitoring data recovery model.

[0089] The streaming network monitoring data recovery model includes a context information extraction module, an event encoding module, and a data recovery module. For a historical measurement event dataset, the input to the context information extraction module is a bipartite graph sequence, and the input to the event encoding module is the historical measurement event data from the Kth round.

[0090] To overcome inherent data sparsity and extract data node-by-node, a graph neural network is introduced to generate source node factor matrices and target node factor matrices. To learn node-level temporal dynamics through the outputs of the graph neural network (i.e., the source node factor matrices and target node factor matrices), a recurrent neural network is used to generate the source node context feature vector A for subsequent matrix filling tasks. ctx and the target node context feature vector B ctx Therefore, the context information extraction module includes a graph neural network and a recurrent neural network. The graph neural network is used to extract features from the bipartite graph sequence, obtaining the source node factor matrix and the target node factor matrix. The recurrent neural network is used to extract features from all source node factor matrices of the bipartite graph sequence, obtaining the source node context feature vector A. ctx Feature extraction is performed on the factor matrices of all target nodes in the bipartite graph sequence to obtain the target node context feature vector B. ctx .

[0091] The event encoding module is an invariant flow event encoder that is insensitive to the input order. It is used to extract features from the historical measurement event data of the Kth round to obtain the source node state vector and the target node state vector.

[0092] The data recovery module is used to analyze the source node's context feature vector A. ctx and source node state vector A curr Target node context feature vector B ctx and the target node state vector B curr Deep feature extraction is performed to obtain the state prediction matrix of the Kth round of historical measurement event data.

[0093] Step 5: Train the streaming network monitoring data recovery model to obtain the target streaming network monitoring data recovery model.

[0094] The streaming network monitoring data recovery model is trained using the Kth round of historical measurement time data and the bipartite graph sequence of each historical measurement event dataset as input and the corresponding target matrix as output. The specific implementation process includes:

[0095] Step 5.1: Use the context information extraction module to extract features from the bipartite graph sequence to obtain the source node context feature vector and the target node context feature vector;

[0096] Step 5.2: Use the event encoding module to extract features from the corresponding Kth round of historical measurement event data to obtain the source node state vector and the target node state vector;

[0097] Step 5.3: Use the data recovery module to perform deep feature extraction on the source node context feature vector and source node state vector, the target node context feature vector and target node state vector to obtain the state prediction matrix of the Kth round of historical measurement event data;

[0098] Step 5.4: Construct a loss function based on the state prediction matrix and the target matrix, and adjust the parameters of the streaming network monitoring data recovery model with the goal of minimizing the loss function to achieve model training.

[0099] In step 5.1, the context information extraction module is used to extract features from the bipartite graph sequence to obtain the source node context feature vector and the target node context feature vector, specifically including:

[0100] Step 5.11: For each bipartite graph in the bipartite graph sequence, randomly generate an initial embedding matrix of source nodes based on the number of source nodes corresponding to the bipartite graph. An initial embedding matrix of target nodes is randomly generated based on the number of target nodes in the bipartite graph. Where |S| is the number of source nodes, |D| is the number of target nodes, dim is the embedding dimension, and R is the real number space.

[0101] The embedding dimension dim is set based on experience; in this embodiment, dim = 32.

[0102] Step 5.12: Initial embedding matrix based on source nodes and the initial embedding matrix of the target node Feature extraction is performed on the bipartite graph to obtain the source node factor matrix and the target node factor matrix.

[0103] In step 5.12, the specific implementation process for obtaining the source node factor matrix and the target node factor matrix includes:

[0104] Step 5.121: For any edge (s,d) in the bipartite graph, take the initial embedding matrix of the source node. The s-th row is used as the initial embedding feature s (0) Take the initial embedding matrix of the target node. The d-th row is used as the initial embedding feature d (0) Where s is the starting node ID of edge (s,d) and d is the ending node ID of edge (s,d).

[0105] Step 5.122: Initial embedding features s of the edge (s,d) of the bipartite graph (0) Feature encoding and feature aggregation are performed to obtain the embedding features of the starting node of each edge; the initial embedding features d of the edge (s,d) of the bipartite graph are also obtained. (0) Feature encoding and feature aggregation are performed to obtain the embedding features of the endpoint node of each edge.

[0106] For any edge (s, d) in a bipartite graph, the message from s to d is defined as: m s→d =f msg (s (0) ,d (0) ,e sd ), where f msg (·) is the feature encoding function, which uses the initial embedded features s. (0) Initial embedded features d (0) The value e of the edge sd (i.e., the measurement result) is used as input for feature encoding to obtain the intermediate feature encoding vector m. s→d f msg (s (0) ,d (0) ,e sd )=W[s (0) ;d (0) ;e sd ]+b), where W is a trainable weight matrix, b is a trainable bias matrix, and [.;.] is a vector concatenation operation.

[0107] Encoding the features of all edges in a bipartite graph yields the corresponding intermediate feature encoding vectors. Features are encoded and propagated along the edges of the bipartite graph. When a node has multiple edges, it will receive multiple features; therefore, feature aggregation is required. Taking the endpoint node d as an example, feature aggregation is achieved using the following formula:

[0108]

[0109] in, Let g() be the set of all nodes pointing to node d, and g() be a trainable linear transformation function. Feature encoding and feature aggregation accurately show how features propagate to the first-order neighborhood; if feature propagation is repeated, higher-order features on the bipartite graph can be explored. In a multi-layer graph neural network, the endpoint node d is recursively updated, specifically by the following formula:

[0110]

[0111] Where, d (L) Let d be the embedding feature of the endpoint node d in the Lth layer. (L-1) The embedding feature of the endpoint node d in the L-1 layer; u (L-1) Let be the embedding feature of node u in the L-1 layer.

[0112] Similarly, for edge (s,d), the embedding features of the starting node s at layer L can be obtained. Repeating steps 5.121 and 5.122, for each edge in the bipartite graph, the embedding features of the starting node and the ending node at layer L can be obtained.

[0113] Step 5.123: Replace the initial embedding matrix of the source nodes with the embedding features of all starting nodes of the bipartite graph at layer L. The corresponding rows are used to obtain the source node factor matrix. The initial embedding matrix of the target node is then replaced with the embedding features of all endpoint nodes in the bipartite graph at layer L. The corresponding rows yield the target node factor matrix.

[0114] Repeat steps 5.11 to 5.12 to obtain the source node factor matrix and target node factor matrix for each bipartite graph.

[0115] Step 5.13: Extract features from the factor matrices of all source nodes in the bipartite graph sequence to obtain the source node context feature vector A. ctx Feature extraction is performed on the factor matrices of all target nodes in the bipartite graph sequence to obtain the target node context feature vector B. ctx The specific formula is as follows:

[0116]

[0117] Among them, S 1 ,S 2 ,…,S (K-2) ,S (K-1) G is a bipartite graph. 1 G 2 ,…,G (K-2) G (K-1) The source node factor matrix, D 1 D 2 ,…,D (K-2) D (K-1) G is a bipartite graph. 1 G 2 ,…,G (K-2) G (K-1) The target node factor matrix, GRU represents a gated recurrent neural network.

[0118] In step 5.2, the event encoding module is used to extract features from the corresponding Kth round of historical measurement event data to obtain the source node state vector and the target node state vector. The specific implementation process includes:

[0119] Step 5.21: Randomly generate the initial learning matrix of the source nodes based on the number of source nodes in the historical measurement event data of the Kth round. An initial learning matrix for target nodes is randomly generated based on the number of target nodes in the historical measurement event data of the Kth round.

[0120] The embedding dimension dim is set based on experience; in this embodiment, dim = 32.

[0121] Step 5.22: Define the source node memory matrix M S and the target node memory matrix M D Source node memory matrix M S The number of rows and columns and the initial learning matrix of the source node The same applies to the target node memory matrix M. D The number of rows and columns and the initial learning matrix of the target node Similarly, when a new round of measurement begins, the source node memory matrix M is... S and the target node memory matrix M D Initialize to 0.

[0122] Step 5.23: For a measurement event e = (p, q, t, v) in the Kth round of historical measurement event data, take the initial learning matrix of the source node. The p-th row is taken as P, and the initial learning matrix of the target node is taken. The qth row is Q; where p is the source node ID, q is the target node ID, t is the timestamp, and v is the measurement result.

[0123] Step 5.23: Memorize the source node matrix M S and the target node memory matrix M D To update, the specific update formula is as follows:

[0124]

[0125] Among them, Updater k (·) represents a single-layer network, Updater k (x)=ReLU(W k x+bk), W k and b k For learnable parameters, ReLU(x) = max(x, 0); the measurement events are encoded as: h = [P; Q; v]; M is the source node memory matrix after the i-th update. S , Let M be the memory matrix of the target node after the i-th update. D , i = 1, 2, ..., M, where the update number M is the number of measurement events in the historical measurement event data of the Kth round.

[0126] Step 5.24: Use the max pooling layer to memorize the updated source node matrix. and target node memory matrix After processing, the source node state vector A is obtained. curr and the target node state vector B curr The specific formula is as follows:

[0127]

[0128] MaxPooling() is the max-pooling layer. To capture representations across nodes, the memory matrix is ​​obtained from the source nodes in the current measurement round. Target node memory matrix It is maintained by max pooling.

[0129] In step 5.3, the data recovery module is used to perform deep feature extraction on the source node context feature vector and source node state vector, the target node context feature vector and target node state vector to obtain the state prediction matrix of the Kth round of historical measurement event data. The specific formula is as follows:

[0130]

[0131] in, Here is the state prediction matrix, and MLP is a multilayer perceptron model;

[0132] In step 5.4, the specific expression for the loss function is as follows:

[0133]

[0134] in, For loss function, Let X be the state prediction matrix. Ω Let |Ω| be the target matrix, and |Ω| be the target matrix X. Ω The number of non-zero elements in the middle, ||·|| F Let λ represent the Frobenius norm, where λ is a constant. In this implementation, λ is set to 0.1. The prediction matrix for the Kth round of historical measurement event data The value at position (i,j), The target matrix X is the historical measurement event data of the Kth round. Ω The value of the position (i,j).

[0135] To measure the accuracy of the streaming network monitoring data recovery model, two metrics are defined (the smaller the better): Normalized Mean Absolute Error (NMAE) and Normalized Root Mean Square Error (NRMSE), with the following formulas:

[0136]

[0137]

[0138] The ResNMMF model of this invention was compared with mathematical low-rank tensor filling algorithms (FaLRTC, HaLRTC, TNCP, CP-als) on different datasets (Seattle, NYC Taxi, GEANT, Abilene), and the comparison results are as follows. Figure 2 As shown. Figure 2 In, FaLRTC stands for Fast Low Rank Tensor Completion (Reference: Liu J, Musialski P, Wonka P, et al. Tensor completion for estimating missing values ​​in visual data [J]. IEEE transactions on pattern analysis and machine intelligence, 2012, 35(1): 208-220.), and HaLRTC stands for High Accuracy Low Rank Tensor Completion (Reference: Liu J, Musialski P, Wonka P,et al.Tensor completion for estimating missing values ​​in visual data[J].IEEE transactions on pattern analysis and machineintelligence,2012,35(1):208-220.), TNCP represents Trace Norm Regularized CANDECOMP / PARAFAC Decomposition (Reference Liu Y, Shang F, Jiao L, et al. Trace normregularized CANDECOMP / PARAFAC decomposition with missing data[J].IEEEtransactions on cybernetics,2014,45(11):2437-2448.), CP-als means CANDECOMP / PARAFAC Decomposition with alternative least square. Figure 2 Figure (a) shows the NMAE and NRMSE indices of the ResNMMF model and the mathematical low-rank tensor padding algorithm on the Abilene dataset at different sampling rates. Figure 2Figure (b) shows the NMAE and NRMSE indices of the ResNMMF model and the mathematical low-rank tensor padding algorithm on the GEANT dataset at different sampling rates. Figure 2 Figure (c) shows the NMAE and NRMSE indices of the ResNMMF model and the mathematical low-rank tensor padding algorithm on the NYCTaxi dataset at different sampling rates. Figure 2 Figure (d) shows the NMAE and NRMSE indices for the ResNMMF model and the mathematical low-rank tensor padding algorithm on the Seattle dataset at different sampling rates. Figure 2 It is evident that the ResNMMF model outperforms the mathematical low-rank tensor imputation algorithm in terms of accuracy. This is because the ResNMMF model employs a neural network to recover missing values. Compared to the lower expressive power of the mathematical low-rank tensor imputation algorithm, the neural network possesses stronger expressive power, resulting in lower values ​​for both error metrics. The datasets used in this invention are: Seattle (representing Seattle road network speeds), NYC Taxi (representing New York City taxis), GEANT, and Abilene. The sources of these four datasets can be found in [link to relevant documentation].

[0139] https: / / github.com / MerrillLi / TensorDataset / releases / tag / v1.

[0140] The ResNMMF model of this invention was compared with neural network tensor filling algorithms (CoSTCo, NTM, NTC) on different datasets (Seattle, NYC Taxi, GEANT, Abilene), and the comparison results are as follows. Figure 3 As shown. Figure 3, CoSTCo represents Convolution Sparse Tensor Completion (Reference: Liu H, Li Y, Tsang M, etal. Costco: A neural tensor completion model for sparse tensors [C] / / Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery&DataMining.2019:324-334.), NTM represents Neural Tensor Model (Reference: Chen H, Li J.Neuraltensor model for learning multi-aspect factors in recommender systems[C] / / International Joint Conference on Artificial Intelligence(IJCAI).2020,2020.), NTC stands for Neural Tensor Completion (reference Xie K, Lu H, Wang X, et al.Neural tensor completion for accurate network monitoring[C] / / IEEE INFOCOM 2020-IEEEConference on Computer Communications. IEEE, 2020:1688-1697.). Figure 3 Figure (a) shows the NMAE and NRMSE indices of the ResNMMF model and the neural network tensor filling algorithm on the Abilene dataset at different sampling rates. Figure 3 Figure (b) shows the NMAE and NRMSE indices of the ResNMMF model and the neural network tensor filling algorithm on the GEANT dataset at different sampling rates. Figure 3 Figure (c) shows the NMAE and NRMSE indices of the ResNMMF model and the neural network tensor filling algorithm on the NYC Taxi dataset at different sampling rates. Figure 3 Figure (d) shows the NMAE and NRMSE indices for the ResNMMF model and the neural network tensor filling algorithm on the Seattle dataset at different sampling rates. Figure 3It is evident that the ResNMMF model outperforms neural network tensor filling algorithms in most cases in terms of accuracy. This is because the ResNMMF model employs a sliding window design, which better utilizes the strong locality of network states, thus performing well at lower sampling rates. However, when the sampling rate is high, the ResNMMF model's accuracy advantage over neural network tensor filling algorithms becomes less significant, and may even be inferior. This is because the event encoding module has limitations in its encoding capabilities, unable to capture a large number of events. Despite this slight disadvantage in accuracy, the ResNMMF model performs real-time filling, possessing unparalleled online running speed and real-time filling characteristics compared to other tensor filling algorithms.

[0141] The ResNMMF model of this invention was compared with the online learning-based tensor filling algorithm STF (reference: Ahn D, Kim S, Kang U. Accurate online tensor factorization for temporal tensor streams with missing values ​​[C] / / Proceedings of the 30th ACM International Conference on Information & Knowledge Management. 2021:2822-2826.) on different datasets (GEANT, Abilene). The comparison results are as follows: Figure 4 As shown. Figure 4 Figure (a) shows the NRMSE index values ​​of the ResNMMF model and STF for different time slots on the Abilene dataset. Figure 4 Figure (b) shows the NRMSE index values ​​of the ResNMMF model and STF for different time slots on the GEANT dataset. Figure 4 Figure (c) shows the NMAE index values ​​of the ResNMMF model and STF for different time slots on the Abilene dataset. Figure 4 Figure (d) shows the NMAE index values ​​of the ResNMMF model and the STF model for different time slots on the GEANT dataset. Figure 4 It can be seen that on the GEANT and Abilene datasets, the ResNMMF model has a certain advantage in terms of accuracy (the lower the better). This is because the ResNMMF model uses a neural network, which has stronger expressive power.

[0142] As shown in Table 1, compared to STF, the ResNMMF model demonstrates faster event encoding module speed and shorter processing time on four publicly available datasets. The figures in Table 1 represent the average time to process one round of historical measurement event data, in milliseconds (ms).

[0143] Table 1 Comparison of ResNMMF model and STF on four datasets

[0144] Seattle 14.2 33.1 NYC Taxi 0.37 4.42 GEANT 0.24 3.47 Abilene 0.14 3.31

[0145] Compared with the online learning-based tensor filling algorithm STF, the ResNMMF model of this invention has better accuracy and faster processing speed in real-time filling scenarios.

[0146] The inference time (training period) of each part of the ResNMMF model of this invention is shown in Table 2:

[0147]

[0148] In Table 2, Dataset represents the dataset, Sampling Rate represents the sampling rate, Total represents the total inference time, and Stage represents the different computational stages of the ResNMMF model: GNN stage (the process of transforming the bipartite graph sequence into the source node factor matrix sequence and the target node factor matrix sequence), RNN stage (converting the factor matrix sequence into A... ctx and B ctx The process), Streaming represents the online processing time of the event encoding module, and Inference represents the processing time of the data recovery module. The unit of time is ms (milliseconds).

[0149] Example 2

[0150] Based on the same concept, embodiments of the present invention also provide a method for recovering streaming network monitoring data, comprising the following steps:

[0151] Step 1: Obtain network measurement event data from the previous K-1 rounds and network measurement event data from the current round;

[0152] Step 2: Construct a bipartite graph for each round of network measurement event data in the first K-1 rounds of network measurement event data;

[0153] Step 3: Invoke the target streaming network monitoring data recovery model, which is trained according to the streaming network monitoring data recovery model training method described above;

[0154] Step 4: Based on the target streaming network monitoring data recovery model, perform data recovery on the current round of network measurement event data to obtain the state prediction matrix of the current round of network measurement event data. The state prediction matrix is ​​the recovered network monitoring data.

[0155] In step 2, the specific implementation process of constructing a bipartite graph for each round of network measurement event data is as follows:

[0156] In each round of network measurement event data, an undirected edge is created between the source node p and the target node q of each measurement event, with the value e of the undirected edge. pq Let v be the measurement result of this measurement event, from which we obtain the bipartite graph G of the measurement event data for this round. (k) This yields a bipartite graph sequence of network measurement event data from the first K-1 rounds, denoted as {G}. 1 G 2 ,…,G (K-2) G (K-1)}

[0157] In step 4, the specific implementation process for data recovery of the current round of network measurement event data is as follows:

[0158] Step 4.1: Use the context information extraction module of the target streaming network monitoring data recovery model to extract features from the bipartite graph sequence in Step 2 to obtain the source node context feature vector and the target node context feature vector;

[0159] Step 4.2: Use the event encoding module of the target streaming network monitoring data recovery model to extract features from the current round of network measurement event data to obtain the source node state vector and the target node state vector;

[0160] Step 4.3: Using the data recovery module of the target streaming network monitoring data recovery model, perform deep feature extraction on the source node context feature vector and source node state vector, the target node context feature vector and target node state vector to obtain the state prediction matrix of the current round of network measurement event data, which is the recovered network monitoring data.

[0161] In step 4.1, the context information extraction module is used to extract features from the bipartite graph sequence, obtaining the source node context feature vector and the target node context feature vector, specifically including:

[0162] Step 4.11: For each bipartite graph in the bipartite graph sequence, randomly generate an initial embedding matrix of source nodes based on the number of source nodes corresponding to the bipartite graph. An initial embedding matrix of target nodes is randomly generated based on the number of target nodes in the bipartite graph. Where |S| is the number of source nodes, |D| is the number of target nodes, dim is the embedding dimension, and R is the real number space.

[0163] Step 4.12: Initial embedding matrix based on source nodes and the initial embedding matrix of the target node Feature extraction is performed on the bipartite graph to obtain the source node factor matrix and the target node factor matrix.

[0164] In step 4.12, the specific implementation process for obtaining the source node factor matrix and the target node factor matrix includes:

[0165] Step 4.121: For any edge (s, d) in the bipartite graph, take the initial embedding matrix of the source node. The s-th row is used as the initial embedding feature s (0) Take the initial embedding matrix of the target node. The d-th row is used as the initial embedding feature d (0) Where s is the starting node ID of edge (s,d) and d is the ending node ID of edge (s,d).

[0166] Step 4.122: Initial embedding features s of the edge (s,d) of the bipartite graph (0) Feature encoding and feature aggregation are performed to obtain the embedding features of the starting node of each edge; the initial embedding features d of the edge (s,d) of the bipartite graph are also obtained. (0) Feature encoding and feature aggregation are performed to obtain the embedding features of the endpoint node of each edge, as shown in formulas (1) and (2).

[0167] Step 4.123: Replace the initial embedding matrix of the source nodes with the embedding features of all starting nodes of the bipartite graph at layer L. The corresponding rows are used to obtain the source node factor matrix. The initial embedding matrix of the target node is then replaced with the embedding features of all endpoint nodes in the bipartite graph at layer L. The corresponding rows yield the target node factor matrix.

[0168] Repeat steps 4.11 to 4.12 to obtain the source node factor matrix and target node factor matrix for each bipartite graph. For each bipartite graph sequence, the source node factor matrix sequence is formed by the source node factor matrices of all bipartite graphs, and the target node factor matrix sequence is formed by the target node factor matrices of all bipartite graphs.

[0169] Step 4.13: Extract features from the factor matrices of all source nodes in the bipartite graph sequence to obtain the source node context feature vector A. ctx Feature extraction is performed on the factor matrices of all target nodes in the bipartite graph sequence to obtain the target node context feature vector B. ctx As shown in formula (3).

[0170] In step 4.2, the event encoding module is used to extract features from the network measurement event data of the current round to obtain the source node state vector and the target node state vector. The specific implementation process includes:

[0171] Step 4.21: Randomly generate the initial learning matrix for the source nodes based on the number of source nodes in the current round of network measurement event data. An initial learning matrix for the target nodes is randomly generated based on the number of target nodes in the current round of network measurement event data.

[0172] Step 4.22: Define the source node memory matrix M S and the target node memory matrix M D Source node memory matrix M S The number of rows and columns and the initial learning matrix of the source node The same applies to the target node memory matrix M. D The number of rows and columns and the initial learning matrix of the target node Similarly, the source node memory matrix M is... S and the target node memory matrix M D Initialize to 0.

[0173] Step 4.23: For a measurement event e = (p, q, t, v) in the current round of network measurement event data, take the initial learning matrix of the source node. The p-th row is taken as P, and the initial learning matrix of the target node is taken. The qth row is Q; where p is the source node ID, q is the target node ID, t is the timestamp, and v is the measurement result.

[0174] Step 4.23: Memorize the source node matrix M S and the target node memory matrix M D The update is performed, and the specific update formula is shown in equation (4).

[0175] Step 4.24: Use the max pooling layer to memorize the updated source node matrix. and target node memory matrix After processing, the source node state vector A is obtained. curr and the target node state vector B curr .

[0176] In step 4.3, the data recovery module is used to perform deep feature extraction on the source node context feature vector and source node state vector, the target node context feature vector and target node state vector to obtain the state prediction matrix of the current round network measurement event data, as shown in equation (6).

[0177] This invention proposes an asynchronous processing approach that combines online processing (context information extraction module) and offline processing (event encoding module), allowing these two parts to be processed separately, thereby improving the model response speed when deployed in measurement scenarios. The event encoding module of this invention is faster than event processing based on recurrent neural networks during inference calculations, and consumes less computational resources and time overhead in model training. At the same time, the event encoding module also possesses set invariance, which is not available in recurrent neural networks (RNNs), ensuring that the calculation results are independent of the order in which events arrive and are only related to the received measurement events.

[0178] Based on the same concept, embodiments of the present invention also provide an electronic device, including: a processor and a memory storing a computer program, wherein the processor is configured to implement the streaming network monitoring data recovery model training method or the streaming network monitoring data recovery method as described above when executing the computer program.

[0179] Although not shown, the electronic device includes a processor that can perform various appropriate operations and processes based on programs and / or data stored in read-only memory (ROM) or loaded from a storage portion into random access memory (RAM). The processor can be a multi-core processor or may include multiple processors. In some embodiments, the processor may include a general-purpose main processor and one or more specialized coprocessors, such as a central processing unit (CPU), graphics processing unit (GPU), neural network processor (NPU), digital signal processor (DSP), etc. Various programs and data required for the operation of the electronic device are also stored in the RAM. The processor, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0180] The processor and memory described above are used together to execute a program stored in the memory. When the program is executed by a computer, it can implement the methods, steps, or functions described in the above embodiments.

[0181] Although not shown, embodiments of the present invention also provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the streaming network monitoring data recovery model training method or the streaming network monitoring data recovery method as described above.

[0182] Storage media in embodiments of the present invention include articles that are permanent or non-permanent, removable or non-removable, and can store information by any method or technology. Examples of storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information that can be accessed by a computing device.

[0183] The above description only discloses specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or modifications that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for training a streaming network monitoring data recovery model, characterized in that, The method includes: Acquire N rounds of historical measurement event data, where each round of historical measurement event data includes n historical measurement events, and each historical measurement event includes source node ID, target node ID, timestamp, and measurement result; A historical measurement event dataset is constructed by selecting K consecutive historical measurement event data from N rounds of historical measurement event data, thus obtaining N-K+1 historical measurement event datasets; A bipartite graph is constructed for the first K-1 rounds of historical measurement event data in each historical measurement event dataset, thereby obtaining a bipartite graph sequence for each historical measurement event dataset; the Kth round of historical measurement event data in each historical measurement event dataset is transformed into a target matrix; A streaming network monitoring data recovery model is constructed, comprising a context information extraction module, an event encoding module, and a data recovery module; the context information extraction module and the event encoding module are respectively connected to the data recovery module. The bipartite graph sequence of each historical measurement event dataset is used as the input to the context information extraction module, the Kth round of historical measurement time data of the corresponding historical measurement event dataset is used as the input to the event encoding module, and the corresponding target matrix is ​​used as the model output. The streaming network monitoring data recovery model is trained to obtain the target streaming network monitoring data recovery model.

2. The training method for a streaming network monitoring data recovery model according to claim 1, characterized in that, For each of the historical measurement event datasets, a bipartite graph is constructed from the first K-1 rounds of historical measurement event data, specifically including: For the k-th round of historical measurement event data, an undirected edge is created between the source node and the target node of each historical measurement event. The value of the undirected edge is the measurement result of the historical measurement event, thus obtaining a bipartite graph of the k-th round of historical measurement event data; where k = 1, 2, ..., K-1.

3. The method for training a streaming network monitoring data recovery model according to claim 1, characterized in that, The specific implementation process of training the streaming network monitoring data recovery model includes: The context information extraction module is used to extract features from the bipartite graph sequence to obtain the source node context feature vector and the target node context feature vector. The event encoding module is used to extract features from the corresponding Kth round of historical measurement event data to obtain the source node state vector and the target node state vector; The data recovery module is used to perform deep feature extraction on the source node context feature vector and source node state vector, the target node context feature vector and target node state vector to obtain the state prediction matrix of the Kth round of historical measurement event data; A loss function is constructed based on the state prediction matrix and the target matrix. The parameters of the streaming network monitoring data recovery model are adjusted with the goal of minimizing the loss function, thereby achieving model training.

4. The method for training a streaming network monitoring data recovery model according to claim 3, characterized in that, The context information extraction module is used to extract features from the bipartite graph sequence to obtain the source node context feature vector and the target node context feature vector, specifically including: For each bipartite graph in the bipartite graph sequence, an initial embedding matrix of source nodes is randomly generated based on the number of source nodes corresponding to the bipartite graph. An initial embedding matrix of target nodes is randomly generated based on the number of target nodes in the bipartite graph. Where |S| is the number of source nodes, |D| is the number of target nodes, dim is the embedding dimension, and R is the real number space; Based on the initial embedding matrix of the source node and the initial embedding matrix of the target node Feature extraction is performed on the bipartite graph to obtain the source node factor matrix and the target node factor matrix; Feature extraction is performed on the factor matrices of all source nodes in the bipartite graph sequence to obtain the source node context feature vector A. ctx Feature extraction is performed on the factor matrices of all target nodes in the bipartite graph sequence to obtain the target node context feature vector B. ctx The specific formula is as follows: THE ctx =GRU({S 1 ,S 2 ,…,S (K-2) ,S (K-1) }) B ctx =GRU({D 1 ,D 2 ,…,D (K-2) ,D (K-1) }) Among them, S 1 ,S 2 ,…,S (K-2) ,S (K-1) G is a bipartite graph. 1 G 2 ,…,G (K-2) G (K-1) The source node factor matrix, D 1 D 2 ,…,D (K-2) D (K-1) G is a bipartite graph. 1 G 2 ,…,G (K-2) G (K-1) The target node factor matrix, GRU represents a recurrent neural network.

5. The method for training a streaming network monitoring data recovery model according to claim 4, characterized in that, Based on the initial embedding matrix of the source node and the initial embedding matrix of the target node Feature extraction is performed on the bipartite graph to obtain the source node factor matrix and the target node factor matrix, specifically including: For any edge (s, d) in a bipartite graph, take the initial embedding matrix of the source node. The s-th row is used as the initial embedding feature s (0) Take the initial embedding matrix of the target node. The d-th row is used as the initial embedding feature d (0) Where s is the starting node ID of edge (s,d), and d is the ending node ID of edge (s,d). The initial embedding features s of the edge (s,d) of the bipartite graph (0) Feature encoding and feature aggregation are performed to obtain the embedding features of the starting node of edge (s,d); the initial embedding features d of edge (s,d) of the bipartite graph are obtained. (0) Feature encoding and feature aggregation are performed to obtain the embedding features of the endpoint node of edge (s,d); subsequently, the embedding features of the starting node and the endpoint node of each edge in the bipartite graph are obtained; the specific formula is as follows: Where, r (L) Let r be the embedding feature of node r in layer L. (L-1) Let g be the embedding feature of node r in the (L-1)th layer, and g() be a trainable linear transformation function. Let r be the set of all nodes that point to node r. Let W be the intermediate feature encoding vector, W be the trainable weight matrix, b be the trainable bias matrix, and u be the intermediate feature encoding vector. (L-1) Let e ​​be the embedding feature of node u in the (L-1)th layer. ur Let (u,r) be the value of the edge. The initial embedding matrix of the source nodes is replaced with the embedding features of all starting nodes in the bipartite graph at layer L. The corresponding rows are used to obtain the source node factor matrix. The initial embedding matrix of the target node is then replaced with the embedding features of all endpoint nodes in the bipartite graph at layer L. The corresponding rows are used to obtain the target node factor matrix.

6. The method for training a streaming network monitoring data recovery model according to claim 3, characterized in that, The event encoding module is used to extract features from the corresponding Kth round of historical measurement event data to obtain the source node state vector and the target node state vector. The specific implementation process includes: An initial learning matrix for source nodes is randomly generated based on the number of source nodes in the historical measurement event data of the Kth round. An initial learning matrix for target nodes is randomly generated based on the number of target nodes in the historical measurement event data of the Kth round. Where |S| is the number of source nodes, |D| is the number of target nodes, dim is the embedding dimension, and R is the real number space; Define the source node memory matrix M S and the target node memory matrix M D Source node memory matrix M S The number of rows and columns and the initial learning matrix of the source node The same applies to the target node memory matrix M. D The number of rows and columns and the initial learning matrix of the target node The same, and the source node memory matrix M S and the target node memory matrix M D Initialize to 0; For the measurement event e = (p, q, t, v) in the historical measurement event data of the Kth round, take the initial learning matrix of the source node. The p-th row is taken as P, and the initial learning matrix of the target node is taken. The q-th row is taken as Q; where p is the source node ID, q is the target node ID, t is the timestamp, and v is the measurement result; Memory matrix M of source nodes S and the target node memory matrix M D To update, the specific update formula is as follows: Among them, Updater k (·) represents a single-layer network, Updater k (x)=ReLU(W k x+b k ), W k and b k For learnable parameters, ReLU(x) = max(x, 0); h = [P; Q; v]; M is the source node memory matrix after the i-th update. S , Let M be the memory matrix of the target node after the i-th update. D i = 1, 2, ..., M, where the update count M is the number of measurement events in the historical measurement event data of the Kth round; The updated source node memory matrix is ​​used with a max pooling layer. and target node memory matrix After processing, the source node state vector A is obtained. curr and the target node state vector B curr The specific formula is as follows: MaxPooling() is the maximum pooling layer.

7. The method for training a streaming network monitoring data recovery model according to claim 3, characterized in that, The specific expression for the loss function is as follows: in, For loss function, Let X be the state prediction matrix. Ω Let |Ω| be the target matrix, and |Ω| be the target matrix X. Ω The number of non-zero elements in the middle, ||·|| F Let λ denote the Frobenius norm, where λ is a constant.

8. A method for recovering streaming network monitoring data, comprising the following steps: Acquire network measurement event data from the previous K-1 rounds and network measurement event data from the current round; Construct a bipartite graph for each round of network measurement event data in the first K-1 rounds of network measurement event data; The target streaming network monitoring data recovery model is invoked, and the target streaming network monitoring data recovery model is trained according to the streaming network monitoring data recovery model training method described in any one of claims 1 to 7; Based on the target streaming network monitoring data recovery model, the current round of network measurement event data is recovered to obtain the state prediction matrix of the current round of network measurement event data. The state prediction matrix is ​​the recovered network monitoring data.

9. An electronic device, characterized in that, The device includes: Memory, used to store computer programs; A processor, configured to implement the streaming network monitoring data recovery model training method as described in any one of claims 1 to 7 or the streaming network monitoring data recovery method as described in claim 8 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that, when executed by a processor, implements the streaming network monitoring data recovery model training method as described in any one of claims 1 to 7 or the streaming network monitoring data recovery method as described in claim 8.

Images