Training method and device of graph learning system, and graph data node classification method

Abstract

The application provides a kind of graph representation learning system training method, the graph learning system includes graph convolution network, it is used to obtain the feature representation of target node according to the adjacency matrix in graph data The feature of neighbor node of target node is aggregated to obtain the feature representation of target node, the method comprises: S1, obtaining original graph data information, the original graph data information contains node actual feature matrix, original adjacency matrix;S2, based on original adjacency matrix, establish the edge between the nodes with the same degree to realize the augmentation processing to original graph data information obtains additional feature graph and its corresponding additional feature graph adjacency matrix;S3, based on original graph data information and additional feature graph adjacency matrix, train graph convolution network to convergence.The application does not need to construct deep structure, avoids the problems such as over-smoothing, over-fitting, over-extrusion or model over-complexity that may occur in constructing deep network, and the random edge method effectively improves the robustness of the model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of graph data processing, in particular to the technical field of graph representation learning, and more particularly to a training method of a graph representation learning system, a graph data node classification method and device. BACKGROUND

[0002] All things in the universe, from stars to atoms and molecules, are in constant contact with other individuals. This mutual relationship between individuals can be represented by a graph. There are a large number of graph structure data in real life, such as Sina microblog, Facebook, Twitter, and other social network data, as well as biochemical molecular structure data, brain network data, citation network data, knowledge graph, etc. Graph is a more complex nonlinear structure than tree structure, which contains a series of vertices and edges. Vertices can represent individuals, and edges can represent the relationship between individuals. At the same time, vertices or edges can also be attached with some information (features) describing themselves. By learning the features of vertices and edges, it can be well served for processing graph data. Neural networks have excellent performance in data feature extraction, but simply converting graphs into Euclidean structure data and processing them with traditional neural network methods will cause the loss of topological structure information. Therefore, a more effective graph structure data processing method, graph neural network (GNN), has emerged and developed into multiple branches, such as graph convolutional network (GCN), graph auto-encoder (GAE), graph generative network (GGN), graph attention network (GAT), etc.

[0003] As an important branch of GNN, GCN has received widespread attention from academia and industry due to its excellent performance and simple network structure. For example, the classic graph node classification system is mainly constructed using GCN. However, GCN still has some defects, which limit its wide application in the field of graph data processing. The main defects of GCN are as follows: GCN realizes information propagation through convolution in graph space domain, and the range of information propagation gradually expands with the stacking of convolution layers. Since the network structure of GCN is simple, when the network depth is shallow, the model is insufficient in feature extraction and expression of input data, that is, there is a problem of insufficient information propagation caused by insufficient feature mining.

[0004] In order to expand the information propagation range of the GCN model and avoid the possible information under-propagation problem, a traditional solution is to solve the information under-propagation problem from the depth of the model, and construct a deep GCN model. However, this solution from the depth will face new difficulties: the GCN model will have phenomena such as gradient disappearance, excessive smoothing of the feature distribution of the node on the graph (that is, information over-propagation), rapid decline or even collapse of network performance, and the like when the depth is increased. In addition, although some researchers have successfully avoided performance collapse while constructing a deep GCN model, the model depth reaches hundreds of layers, the structure is too complex, the calculation amount is too large in actual application, and the hardware requirement is high.

[0005] The information transmission mechanism of the GCN model is achieved through neighborhood node information aggregation and updating, that is, the feature expression of the target node is obtained by aggregating the neighbor node information. The GCN model obtains the feature representation of the node as a graph representation learning model, which explicitly extracts the feature information of the node on the graph and implicitly encodes the topological connection information of the graph by limiting the information propagation between connected neighbors. Such information propagation mode makes the GCN perform Laplace smoothing on the center node and its neighbors based on topological connectivity. Due to the particularity of non-Euclidean structure data, although the two nodes on the graph are adjacent in the connected path, their features are not necessarily similar. In this case, the information transmission mode of the GCN has the following problems: 1) there may be information loss and noise, and the signal-to-noise ratio of the message flow is low; 2) too much attention is paid to the nodes within the neighborhood, and the feature extraction is insufficient; 3) there is a lack of explicit use of topological structure information. Therefore, it is impossible to solve various problems of the GCN model by starting from the model depth. We need a more expressive graph representation learning model with a larger feature extraction area and more diverse extractable features to obtain better graph node feature representation and achieve more effective graph node classification, graph classification, edge prediction and the like. SUMMARY

[0006] The purpose of the present application is to overcome the defects of the prior art, and provide a new training method of a graph representation learning system, a graph data node classification method and device.

[0007] According to a first aspect of the present application, a training method of a graph representation learning system is provided, the graph learning system comprising a graph convolution network for aggregating features of neighbor nodes of a target node according to an adjacency matrix in graph data to obtain a feature representation of the target node, the method comprising: S1, obtaining original graph data information, the original graph data information comprising a node actual feature matrix and an original adjacency matrix; S2, based on the original adjacency matrix, establishing edges between nodes with the same degree to realize augmentation processing of the original graph data information to obtain an additional feature graph and a corresponding additional feature graph adjacency matrix; and S3, training the graph convolution network based on the original graph data information and the additional feature graph adjacency matrix until convergence.

[0008] Preferably, the step S2 comprises: S21, establishing edges between nodes with the same degree based on the original adjacency matrix to achieve augmentation processing of the original graph data information; S22, obtaining the degree-equivalent adjacency matrix based on the graph data information after the augmentation processing; S23, obtaining the additional feature graph adjacency matrix after sparsifying the degree-equivalent adjacency matrix.

[0009] Preferably, the degree-equivalent adjacency matrix is:

[0010]

[0011] wherein D oh is a one-hot encoding representation matrix of the node degree value in the original graph data information, and N represents the number of nodes in the original graph data information, and M is the maximum value of the node degree in the original graph data information, is the transpose matrix of D oh .

[0012] Preferably, the step S23 comprises: S231, obtaining a degree set of candidate vertices corresponding to a connected node set in the additional feature graph based on the degree-equivalent adjacency matrix; S232, sampling multiple times from a 0-1 uniform distribution to obtain a candidate probability corresponding to each degree in the degree set of candidate vertices, and selecting a degree corresponding to a candidate probability greater than a candidate vertex degree probability threshold as a vertex degree set after sparsification, wherein the candidate vertex degree probability threshold is an adjustable hyperparameter; S233, taking edges in the augmented graph data corresponding to degrees in the vertex degree set after sparsification as a candidate edge set, sampling multiple times from a 0-1 uniform distribution to obtain a candidate probability corresponding to each edge in the candidate edge set, and selecting an edge corresponding to a candidate probability greater than a candidate edge probability threshold as a final edge set, wherein the candidate edge probability threshold is an adjustable hyperparameter; S234, based on the final candidate edge set in step S233, deleting edges in the augmented graph data that do not belong to the candidate edge set to achieve graph data updating, and taking the adjacency matrix corresponding to the updated graph data as the additional feature graph adjacency matrix.

[0013] Preferably, the graph neural network is trained to convergence with the node actual feature matrix and the additional feature graph adjacency matrix as inputs and the node updated feature matrix obtained based on the forward propagation process as output, wherein the forward propagation process is:

[0014]

[0015] wherein Z represents the node updated feature matrix, X represents the node feature matrix input into the graph neural network, represents the additional feature graph adjacency matrix A strwherein L represents a Laplacian matrix of the original graph, W represents parameters of the graph convolution network, and σ(*) represents an activation function

[0016]

[0017]

[0018] wherein I N represents a unit matrix and represents an additional feature graph adjacency matrix A str corresponding to a degree matrix and

[0019] Preferably, in the step S3, the node actual feature matrix and the additional feature graph adjacency matrix are input, a forward propagation process is performed based on a preset feature integration strategy, and a node updated feature matrix is obtained as output, so as to train the graph neural network to convergence. The preset feature integration strategy is:

[0020] a pre-integration strategy, and a corresponding forward propagation process is:

[0021]

[0022] or a post-integration strategy, and a corresponding forward propagation process is:

[0023]

[0024] or a double integration strategy, and a corresponding forward propagation process is:

[0025]

[0026] wherein Z represents a node updated feature matrix, X represents a node feature matrix input into the graph neural network, represents a Laplacian matrix based on the original adjacency matrix and the additional feature graph adjacency matrix A Str F represents a Laplacian matrix based on the additional feature graph adjacency matrix A str represents a Laplacian matrix based on the original adjacency matrix A, W represents parameters of the graph convolution network, and σ(*) represents an activation function.

[0027] Preferably,

[0028] A F =A str ∪A

[0029] ​​​

[0030]

[0031] Among them, I N Describe the identity matrix and Represents the relationship between matrix A and matrix A F The corresponding degree matrix and

[0032]

[0033]

[0034] in, The adjacency matrix A represents the additional feature graph. str The corresponding degree matrix and

[0035]

[0036]

[0037] in, Represents the original adjacency matrix The corresponding degree matrix and

[0038] Preferably, the graph convolutional network includes one or two graph convolutional layers.

[0039] According to a second aspect of the present invention, a graph data node classification method is provided, the method comprising: T1, acquiring graph data to be classified; T2, acquiring feature representations of nodes in the graph data using a graph representation learning system trained as described in the first aspect of the present invention; and T3, classifying nodes based on the node feature representations acquired in step T2.

[0040] According to a third aspect of the present invention, a graph representation learning system based on the method described in the first aspect of the present invention is provided, for processing graph data to obtain feature representations of each node in the graph data, the system comprising: an additional feature map generation module, configured to establish edges between nodes of the same degree in the graph data to be processed to achieve augmentation processing of the graph data, and obtain an additional feature map and its corresponding additional feature map adjacency matrix; and a graph convolution module, configured to generate feature representations of nodes in the graph data based on the additional feature map adjacency matrix.

[0041] According to a fourth aspect of the present invention, a graph data node classification apparatus is provided, characterized in that the apparatus comprises: a graph representation learning system as described in the third aspect of the present invention, for acquiring feature representations of graph data nodes; and a classifier for classifying nodes based on the feature representations of the nodes.

[0042] Compared with the prior art, the advantages of the present application are that the graph representation learning model of the present application does not need to construct a deep structure, avoiding problems such as over-smoothing, over-fitting, over-compression or model over-complexity that may occur in constructing a deep network, and the randomized edge method effectively improves the robustness of the model; based on the additional feature map obtained by the edge augmentation processing, different feature integration strategies are proposed, and for the previous scheme without augmented data, the graph representation learning model or system constructed based on the above strategies can extract better graph node feature representation performance. BRIEF DESCRIPTION OF DRAWINGS

[0043] The embodiments of the present application are further described below with reference to the accompanying drawings, in which:

[0044] Figure 1 The flowchart of the training method of the graph representation learning system according to the embodiments of the present application is shown.

[0045] Figure 2 The process diagram of obtaining an additional feature map based on the graph data augmentation processing according to the embodiments of the present application is shown. DETAILED DESCRIPTION

[0046] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application is further described in detail below through specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0047] It should be noted that in the description of the embodiments in the present application, i and j are repeatedly used, and i and j are not used to refer to a specific variable, but are used to refer to a general variable in different formulas for the convenience of description.

[0048] As mentioned in the background, the information transmission mechanism of the GCN model in the prior art has the following problems: 1) there may be information loss and noise, and the signal-to-noise ratio of the message flow is low; 2) too much attention is paid to the nodes in the neighborhood, and the feature extraction is insufficient; 3) there is a lack of explicit use of topological structure information. Therefore, in order to improve the insufficient feature mining problem of the GCN model and construct a more effective graph data related application system based on the GCN model, the GCN model needs to have robust learning ability of a wide range of features, while taking into account the locality and globality of feature extraction, i.e. an information transmission GCN graph representation learning model with more expressive power, larger feature extraction area and more diverse extractable features is needed. The solution idea in the prior art mainly improves from a single angle, which cannot comprehensively solve the above problems.

[0049] The GCN realizes information propagation through convolution on a graph space domain. The so-called convolution on a space domain is realized through neighborhood node information aggregation and updating, that is, the feature expression of a target node is obtained by aggregating neighbor node information. A key factor of the convolution process is the initial receptive field of a node, that is, the selection of neighbor nodes. With the deepening of the convolution layer, the receptive field of each node gradually expands, and the perceivable features gradually increase. Therefore, the node feature expression obtained by the model is directly related to the receptive field of the node that can be obtained in the convolution. Under the prior art, the initial receptive field range of each target node is directly determined by the adjacency matrix corresponding to the graph data, and the adjacency matrix of the graph data directly comes from the original topological structure of the graph data and has not been augmented. This limits the expansion of the receptive field of the node in the convolution operation, resulting in insufficient effective feature mining of the node, a small receptive field, and insufficient utilization of the structure, which further limits the effect of applications such as graph node classification, graph classification, and edge prediction based on node features. Therefore, the inventors start from the perspective of improving the receptive field based on the characteristics of the graph data and propose a GCN model based on an additional feature graph to solve the above problems. The so-called additional feature graph is obtained by augmenting the original graph using degree features. The GCN model is trained based on the adjacency matrix of the additional feature graph. Since the additional feature graph is augmented data of the original graph, the receptive field of the node is expanded, and the graph representation learning model obtained by training can be more robust and have a wider information propagation range without constructing a complex deep structure, thereby avoiding the problem of insufficient information propagation.

[0050] According to one embodiment of the present application, the present application proposes a training method of a graph representation learning system based on an additional feature graph, as shown in Figure 1 The training method of the present application comprises: S1, obtaining original graph data information, wherein the original graph data information comprises a node actual feature matrix and an original adjacency matrix; S2, based on the original adjacency matrix, establishing an edge between nodes with the same degree to realize augmentation processing of the original graph data information to obtain an additional feature graph and an additional feature graph adjacency matrix thereof; and S3, training a graph convolution network based on the original graph data information and the additional feature graph adjacency matrix until convergence.

[0051] In order to better understand the present application, the graph data augmentation processing process and the training process of the graph representation learning system are described in detail below based on the drawings and embodiments.

[0052] I. Graph data augmentation processing

[0053] The graph structure data includes but is not limited to images, texts, networks and the like, and basic elements in a certain granularity in selected data are taken as nodes (such as pixels, instances and the like in images, words, phrases and the like in texts, nodes, node groups and the like of a network), the adjacency relationship between nodes is defined based on distance, and the node features and the topological connection features of the graph can be extracted. For the GCN model, the input is the node information of the graph data and the adjacency matrix corresponding to the graph data. The GCN model updates the node information by aggregating the neighbor node information layer by layer to obtain the feature representation of the target node. The application obtains the additional feature graph and the corresponding adjacency matrix through the augmentation processing of the graph data, and takes the additional feature graph and the corresponding adjacency matrix as the input of the model. Since the additional feature graph is obtained by augmenting the graph topological structure using the degree feature, the receptive field of the node is expanded, so that the graph representation learning model trained based on the additional feature graph can not only learn local features, but also learn global features.

[0054] As described in the foregoing embodiments, the additional feature graph is obtained by augmenting the graph topological structure using the degree feature, and the degree feature is a manifestation of the node topological connection structure in the graph data. Nodes with the same degree have the same first-order topological structure. In the application, the long-tail phenomenon of the node degree distribution can be changed by establishing edges between nodes with the same degree, the average path length between nodes can be significantly shortened, the efficiency of feature mining of the graph data can be improved, and the feature mining of various graph data can be realized to obtain the node feature representation. Specifically, the application expands the definition range of the node neighbor based on the degree feature of the node, allows the nodes with the same degree to establish edges, and defines the degree-adjacency matrix A deg The following formula

[0055]

[0056] wherein D oh is a one-hot encoding representation matrix of the node degree value in the original graph data information, and N represents the number of nodes in the original graph data information, and M is the maximum value of the node degree in the original graph data information, is the transpose matrix of D oh . Since the number of edges in the augmented graph data is increased, the range of neighbor nodes of each node is expanded, and the degree-adjacency matrix A deg is directly taken as the adjacency matrix of the additional feature graph and the input of the GCN model, more features can be obviously mined.

[0057] According to one embodiment of the application, a random generation method based on the degree feature is used for augmentation processing of the data. Specifically, part of the node degree values are randomly selected from the node degree distribution, and nodes with the corresponding degree values are taken as candidate nodes; then edges are generated by randomly selecting nodes with the same degree from the candidate nodes to complete the augmentation processing.

[0058] Furthermore, it should be noted that real-world networks often exhibit sparse connectivity, and their corresponding graph data also possesses sparsity. To better ensure the sparsity of node connections and the scalability of graph representation learning models, this invention also proposes a scheme to further sparsify the degree-equal adjacency matrix to obtain an additional feature graph adjacency matrix.

[0059] According to an embodiment of the present invention, the present invention performs further sparsification processing on the equi-adjacency matrix through the following steps: S231, obtaining the degree set of candidate vertices corresponding to the connected node set in the additional feature graph based on the degree equi-adjacency matrix; S232, performing multiple samplings from a 0-1 uniform distribution to obtain the candidate probability corresponding to each degree in the candidate vertex degree set, and selecting the degrees whose corresponding candidate probabilities are greater than the candidate vertex degree probability threshold as the sparsified vertex degree set, wherein the candidate vertex degree probability threshold is an adjustable hyperparameter; S233, using the sparsified vertex degree set... The edges of the vertices corresponding to the degrees in the augmented graph data are the candidate edge set. Multiple sampling is performed from a uniform distribution of 0-1 to obtain the candidate probability of each edge in the candidate edge set, and the edges whose corresponding candidate probabilities are greater than the candidate edge probability threshold are selected as the final edge set. The candidate edge probability threshold is an adjustable hyperparameter. S234. Based on the final candidate edge set in step S233, the connected edges in the augmented graph data that do not belong to the candidate edge set are deleted to update the graph data, and the adjacency matrix corresponding to the updated graph data is used as the adjacency matrix of the additional feature graph.

[0060] Suppose the input graph is G = (V, E), where V represents the set of nodes in graph G, E represents the set of edges in graph G, and A represents the adjacency matrix of graph G. The set of node degrees in the graph is d. Let the additional feature graph be G. str =(V,E) str The set of node degrees is d. str The candidate edge set E for the additional feature map is established based on the degree-equal adjacency matrix. str′ E str′ The set of candidate vertex degrees is d str′ , Assumption Degree d n Node pairs (v i ,v j The probability P of establishing an edge link d n ∈d str The probability of ′ is P degree ; {v i ,v j}∈E str The probability is P link| ,but

[0061] P link =P link|degree P degree

[0062] Where P link|degree ~U(0,1), P degree ~U(0,1).

[0063] According to one embodiment of the present invention, firstly, by sampling probability P degree The range of candidate nodes is narrowed down, and then node pairs are selected from the nodes that meet the degree requirement to establish edges. The purpose of this is to reduce the search space for node pairs and shorten the program execution time. Let P degree The reference threshold is the candidate vertex degree probability threshold. P link| The reference threshold is the candidate edge probability threshold. The sparsification process is as follows: |d is performed through a uniform distribution between 0 and 1. str′ |Sampling can yield set d str′ The probability vector P is composed of the candidate probabilities corresponding to each degree. degree , For set d str′ The candidate probability corresponding to the i-th degree value. When When, retain set d str′ The degree d in i The reduced vertex degree set is The corresponding candidate edge set is Then proceed from a uniform distribution between 0 and 1 Subsequent sampling yields a set. The vector P composed of the candidate probabilities of each corresponding edge link| , Represents a set The candidate probability corresponding to the i-th edge in the equation is when At that time, the corresponding edges are retained, and finally the edge set E of the additional feature map is obtained. str Based on this edge set, the adjacency matrix corresponding to the additional feature map is obtained as A. str .

[0064] According to one example of the invention, such as Figure 2As shown, one graph data with 10 nodes, its corresponding adjacency matrix is A, wherein the nodes with degree 1 are node 4 and node 10, the nodes with degree 2 are node 1, node 2, node 7 and node 9, the nodes with degree 3 are node 5 and node 8, the nodes with degree 4 are node 3 and node 6, and the edge pair between the node pair (node 1, node 7) is established to obtain the adjacency matrix A of the graph data after the augmentation processing str Compared with the original adjacency matrix A, the neighbor node range of node 1 and node 7 is obviously expanded.

[0065] The additional feature graph adjacency matrix obtained after the sparse processing is used for feature mining in the GCN model, which can not only expand the node receptive field and sufficiently mine the effective features of the nodes, but also be closer to the real network and the model has better expansibility.

[0066] II. Training of the graph representation learning system

[0067] The GCN model is a model based on a graph convolution layer, and the GCN model used in the graph representation learning system in the application can be one or more graph convolution layers. The working principle of each graph convolution layer is consistent, that is, the node features (or the node feature representation output by the previous layer) are accepted to perform linear transformation and activation function activation to output the updated feature representation of the node, and the updated feature representation of the output node is aggregated by an aggregation function to aggregate the neighbor nodes of the target node. It can be simply understood that the neighbor node features of the target node are aggregated through the adjacency matrix set.

[0068] Taking the graph representation learning system including only one graph convolution layer as an example, the node feature matrix (denoted as X) and the additional feature graph adjacency matrix A str The updated node feature matrix obtained through the forward propagation is:

[0069]

[0070]

[0071]

[0072] Wherein, Z represents the updated node feature matrix, represents the Laplacian matrix based on the additional feature graph adjacency matrix A str , W represents the parameters of the graph convolution layer, and N represents the unit matrix and represents the additional feature graph adjacency matrix A str , and Since the linear transformation in the forward propagation process of the graph convolution layer and the knowledge related to the activation function are known to those skilled in the art, they are not described in detail in the present application, and are simply described as sigma (*).

[0073] In order to better perform feature mining, the present application also proposes a variety of feature integration strategies based on additional feature maps to obtain better node feature representations. According to one embodiment of the present application, the present application provides three feature integration strategies, which are respectively:

[0074] Mining pre-integration strategy, with node feature matrix and mixed feature map A F The forward propagation process as input is:

[0075]

[0076] A F = A str ∪ A

[0077]

[0078]

[0079] Where I N represents the unit matrix and represents the degree matrix corresponding to the matrix A F , and

[0080] Or, the post-mining integration strategy, with node feature matrix, additional feature map A str , and original input graph A as input, the forward propagation process is:

[0081]

[0082]

[0083]

[0084]

[0085]

[0086] Wherein, represents the original adjacency matrix corresponding to the degree matrix and

[0087] Or, the double integration strategy, using the above two strategies at the same time, and superimposing the features corresponding to the two branch strategies, one branch using the mixed feature map to calculate, and the other branch using the original graph to calculate, and the corresponding forward propagation process is:

[0088]

[0089] Similarly, taking two-layer graph convolutional layer as an example:

[0090] Without using the integration strategy, only the node feature matrix and the additional feature graph adjacency matrix A str The forward propagation process as input is:

[0091]

[0092] Wherein, W (0) , W (1) Respectively represent the first graph convolutional layer and the second graph convolutional layer trainable parameters;

[0093] Using the mining pre-integration strategy, the node feature matrix and the mixed feature graph A F The forward propagation process as input is:

[0094]

[0095] Using the mining post-integration strategy, the node feature matrix, the additional feature graph A str , the original input graph A as input, the forward propagation process is:

[0096]

[0097] Using the double integration strategy, the above two integration strategies are used simultaneously, and the forward propagation process is:

[0098]

[0099] Compared with the prior art, the graph representation learning model of the application does not need to construct deep structure, avoids the problems such as over-smoothing, over-fitting or over-extrusion that may occur in constructing deep network, and the random edge method effectively improves the robustness of the model. Based on the additional feature graph obtained by the edge augmentation processing, different feature integration strategies are proposed. For the previous scheme without augmented data, the graph representation learning model or system constructed based on the above strategies can effectively expand the node receptive field and obtain larger range and more diverse feature extraction.

[0100] After training the GCN model based on different strategies, a more robust graph representation learning model can be obtained, which can be used for graph node classification, graph classification, edge prediction tasks, etc. For the graph node classification task, a classifier can be connected at the output of the GCN model trained based on the application to construct a graph node classification device, which can be applied to recommendation systems, social network monitoring, anomaly detection, drug analysis and other scenarios. For the graph classification task, a pooling layer (such as average pooling, maximum value pooling) can be used at the output of the GCN model trained based on the application to extract graph-level representations, and then sent to an update module to obtain graph-level class updates; for the edge prediction task, the node feature matrix can be multiplied with its transpose matrix at the output of the GCN model trained based on the application to obtain the similarity evaluation between nodes, and then based on a threshold or an update function, different similarities are classified into two categories to obtain the edge prediction result. Since the technology of constructing other application task models or systems based on graph representation learning models is known in the art, it is not described in detail here.

[0101] To verify the effect of the application, the inventors conducted node classification verification on the Cora, Citeseer, Pubmed citation network data sets using the GCN model, ResGCN, SGC, GraphSAGE, GAT, APPNP, MixHop, PNA, DEMO-Net, and the node classification model using the additional feature graph model (AFG) proposed in the application. The accuracy of different models was compared, and the results are shown in Table 1. As can be seen from Table 1, the node classification effect of the node classification model of the model trained based on the application is obviously better than that of the original model without using the additional feature graph, and the original model surpasses other GCN improved models after using the additional feature graph.

[0102] Table 1

[0103]

[0104]

[0105] It should be noted that although the above describes each step in a specific order, it does not mean that each step must be performed in the above specific order, and in fact, some of these steps can be performed concurrently or even in a changed order, as long as the desired function can be achieved.

[0106] The application can be a system, a method and / or a computer program product. The computer program product can include a computer readable storage medium having computer readable program instructions loaded thereon for causing a processor to implement various aspects of the application.

[0107] A computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or raised structures in a groove having instructions recorded thereon, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se, such as radio waves or other freely propagating electromagnetic waves.

[0108] Embodiments of the application have been described above, with the understanding that these embodiments are exemplary only, and are not restrictive, in terms of the scope of the embodiments disclosed. Many modifications and variations of the described embodiments are possible, in light of the above teachings, without departing from the scope and spirit of the described embodiments. The choice of words in this document is intended to best explain the principles of the embodiments, practical application, or technical improvement in the art, or to enable others skilled in the art to utilize the embodiments disclosed herein.

Claims

1. A training method for a graph representation learning system, the graph representation learning system comprising a graph convolutional network, which is used to aggregate the features of the neighboring nodes of a target node based on the adjacency matrix in graph data to obtain a feature representation of the target node, characterized in that, The method includes: S1. Obtain the original graph data information, which includes the actual feature matrix of nodes and the original adjacency matrix; S2. Based on the original adjacency matrix, establish edges between nodes of the same degree to augment the original graph data and obtain an additional feature graph and its corresponding adjacency matrix; wherein, S2 includes: S21. Based on the original adjacency matrix, establish edges between nodes of the same degree to achieve augmentation processing of the original graph data information; S22. Obtain the corresponding degree-equal adjacency matrix based on the augmented graph data information; the degree-equal adjacency matrix is: in, It is the one-hot encoded representation matrix of the node degree values ​​in the original graph data information and , This indicates the number of nodes in the original graph data. The maximum degree of a node in the original graph data. yes The transpose of the matrix; S23. After sparsifying the adjacency matrix of equal degree, an additional feature map adjacency matrix is ​​obtained. S3. Train the graph convolutional network until convergence based on the original graph data information and the adjacency matrix of the additional feature graph.

2. The method according to claim 1, characterized in that, S23 includes: S231. Obtain the degree set of candidate vertices corresponding to the connected node set in the additional feature graph based on the degree equal adjacency matrix; S232. Multiple samplings are performed from a 0-1 uniform distribution to obtain the candidate probability corresponding to each degree in the candidate vertex degree set, and the degrees whose corresponding candidate probabilities are greater than the candidate vertex degree probability threshold are selected as the sparsed vertex degree set, wherein the candidate vertex degree probability threshold is an adjustable hyperparameter. S233. The edges of vertices corresponding to the degrees in the sparsed vertex degree set in the augmented graph data are used as candidate edge sets. Multiple sampling is performed from a 0-1 uniform distribution to obtain the candidate probability of each edge in the candidate edge set. The edges whose corresponding candidate probabilities are greater than the candidate edge probability threshold are selected as the final edge set. The candidate edge probability threshold is an adjustable hyperparameter. S234. Based on the final candidate edge set in step S233, delete the edges in the augmented graph data that do not belong to the candidate edge set to update the graph data, and use the adjacency matrix corresponding to the updated graph data as the adjacency matrix of the additional feature graph.

3. The method according to claim 2, characterized in that, In S3, the graph neural network is trained until convergence using the actual feature matrix of the nodes and the adjacency matrix of the additional feature map as input, and the node updated feature matrix obtained based on the forward propagation process as output. The forward propagation process is as follows: in, This indicates that the node updates the feature matrix. This represents the node feature matrix of the input graph neural network. Representation based on the adjacency matrix of the additional feature map The Laplace matrix, This represents the parameters of a graph convolutional network. This represents the activation function.

4. The method according to claim 3, characterized in that, in, Describe the identity matrix and , Represents the adjacency matrix of the additional feature graph The corresponding degree matrix and .

5. The method according to claim 4, characterized in that, In S3, the graph neural network is trained until convergence by taking the actual feature matrix of the nodes, the original adjacency matrix, and the adjacency matrix of the additional feature map as inputs, and the node updated feature matrix obtained by the forward propagation process based on a preset feature integration strategy as output. The preset feature integration strategy is as follows: The forward propagation process for mining pre-integration strategies is as follows: Alternatively, the post-mining ensemble strategy has the following forward propagation process: Alternatively, the dual integration strategy has the following forward propagation process: in, This indicates that the node updates the feature matrix. This represents the node feature matrix of the input graph neural network. Representation based on the original adjacency matrix and the adjacency matrix of the additional feature map AND operation matrix The Laplace matrix, Representation based on the adjacency matrix of the additional feature map The Laplace matrix, Representation based on the original adjacency matrix The Laplace matrix, This represents the parameters of a graph convolutional network. This represents the activation function.

6. The method according to claim 5, characterized in that, in, Describe the identity matrix and , Representation and matrix The corresponding degree matrix and ; in, Represents the adjacency matrix of the additional feature graph The corresponding degree matrix and ; in, Represents the original adjacency matrix The corresponding degree matrix and .

7. The method according to any one of claims 1-6, characterized in that, The graph convolutional network includes one or two graph convolutional layers.

8. A method for classifying graph data nodes, characterized in that, The method includes: T1. Obtain the graph data for node classification; T2. A graph representation learning system trained using any one of the methods described in claims 1-7 obtains feature representations of nodes in graph data; T3. Classify the nodes based on the node feature representations obtained in step T2.

9. A graph representation learning system based on the method of any one of claims 1-7, used to process graph data to obtain feature representations of each node in the graph data, characterized in that, The system includes: The additional feature graph feature matrix generation module is used to establish edges between nodes with the same degree in the graph data to be processed in order to realize the augmentation processing of the graph data, and obtain the corresponding additional feature graph adjacency matrix based on the augmented graph data. The graph convolution module is used to generate feature representations of nodes in graph data based on the adjacency matrix of the additional feature graph.

10. A graph data node classification device, characterized in that, The device includes: The graph representation learning system as described in claim 9 is used to acquire feature representations of graph data nodes; A classifier is used to classify nodes based on their feature representations.

11. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the method according to any one of claims 1 to 8.

12. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the electronic device to perform the steps of the method as described in any one of claims 1 to 8.

Images