Data processing method, system and equipment based on graph neural network compression and medium

Abstract

The invention relates to a data processing method, system and device based on graph neural network compression and a medium, and the method comprises the following steps: obtaining and preprocessing a historical data set, and obtaining a graph structure data set; based on the graph structure data set, adopting a mixed training method to carry out accelerated training on a pre-constructed graph neural network model to obtain a trained graph neural network model; and processing the target graph structure by using the trained graph neural network model to obtain a data processing result. According to the method, a pre-constructed graph neural network model is trained by adopting a mixed training method, and accelerated training is performed on the graph neural network model based on a hierarchical compression method, so that the graph neural network model is quickly converged; and then, an uncompressed conventional training mode is switched to carry out fine tuning on parameters of the graph neural network model, so that higher convergence speed and optimal accuracy are realized, and the method can be widely applied to the field of big data processing.

Description

Technical Field

[0001] The present invention belongs to the field of big data processing, and specifically relates to a data processing method, system, device and medium based on graph neural network compression. Background Art

[0002] As the core technology for processing graph-structured data, graph neural network (GNN) has demonstrated significant advantages in social network analysis, recommendation systems, biomolecular modeling, natural language processing and other fields in recent years.

[0003] The core operations of GNNs include propagation and transformation. Propagation aggregates information about neighboring nodes, while transformation performs nonlinear transformations on node features through neural networks, thereby capturing complex relationships within the graph structure. However, with the explosive growth of graph data (such as trillion-edge social graphs and biomolecular networks), GNN training faces severe efficiency challenges.

[0004] Existing GNN frameworks (such as PyG and DGL) mainly improve efficiency through hardware acceleration (such as cuBLAS / cuSparse libraries) and distributed training, but still have the following shortcomings:

[0005] 1) Computation and storage coupling: The intermediate results of propagation operations (such as edge features in GAS) occupy a large amount of memory, which can easily cause out-of-memory (OOM) problems, especially in large-scale graphs.

[0006] 2) Redundant calculations are not fully utilized: The repeated calculations of transformation operations are not identified, resulting in low utilization of GPU computing power.

[0007] 3) It is difficult to balance accuracy and efficiency: Although methods such as sampling and quantization improve speed, they may sacrifice model accuracy. Their applicability is limited, especially in scenarios that require high-fidelity features (such as biomedical graph analysis). Summary of the Invention

[0008] In response to the above problems, the purpose of the present invention is to provide a data processing method, system, device and medium based on graph neural network compression, which accelerates the propagation and conversion operations in the graph neural network based on layered compression, thereby improving data processing efficiency.

[0009] To achieve the above object, the present invention adopts the following technical solutions:

[0010] In a first aspect, the present invention provides a data processing method based on graph neural network compression, comprising the following steps:

[0011] Obtain historical data sets and preprocess them to obtain graph structure data sets;

[0012] Based on the graph structure dataset, a hybrid training method is used to accelerate the training of the pre-built graph neural network model to obtain a trained graph neural network model;

[0013] The trained graph neural network model is used to process the target graph structure to obtain the data processing results.

[0014] Furthermore, based on the graph structure dataset, a hybrid training method is used to accelerate the training of the pre-built graph neural network model to obtain a trained graph neural network model, including:

[0015] Compress the given graph structure and apply the generated compressed graph structure to the propagation operator of the graph neural network model to obtain a compression-based propagation operator;

[0016] Compress the node features and apply the generated compressed features to the transformation operator of the graph neural network model to obtain a compression-based transformation operator;

[0017] Arrange the obtained compression-based propagation operators and conversion operators according to actual needs to obtain a graph neural network model that has been preliminarily trained.

[0018] The conventional training mode is used to fine-tune the graph neural network model that has completed the preliminary training to obtain a trained graph neural network model.

[0019] Furthermore, the given graph structure is compressed, and the generated compressed graph structure is applied to the propagation operator of the graph neural network model to obtain the compression-based propagation operator, including:

[0020] The given graph structure is compressed using the CompressGraph method to generate a preliminary compressed graph, which includes original nodes and virtual nodes.

[0021] A double-layer filtering method is used to filter virtual vertices and edges with low performance contribution in the preliminary compression graph to obtain a filtered compression graph;

[0022] A multi-step propagation engine is used to apply the propagation operator of the graph neural network model on the filtered compressed graph to obtain the compression-based propagation operator.

[0023] Furthermore, the double-layer filtering method is used to filter virtual vertices and edges with low performance contribution in the preliminary compression graph to obtain a filtered compression graph, including:

[0024] Perform contribution-based filtering on the virtual nodes in the preliminary compression graph according to a preset contribution threshold until there are no virtual nodes in the compression graph whose contribution value is less than the contribution threshold;

[0025] Starting from the maximum depth of the compressed graph, all virtual nodes are deleted layer by layer until the maximum depth of the compressed graph reaches a preset depth threshold.

[0026] Furthermore, the multi-step propagation engine is used to apply the propagation operator of the graph neural network model on the filtered compressed graph to obtain a compression-based propagation operator, including:

[0027] Determine the depth of each node and edge on the filtered compressed graph according to preset rules;

[0028] According to the depth of each node and edge on the compressed graph, nodes and edges with the same depth are divided into the same subgraph, and the number of all subgraphs is recorded as N;

[0029] Based on the depth size, the propagation operator is applied to the subgraphs with a depth of 1 to N in sequence to propagate the features.

[0030] Furthermore, compressing the node features and applying the generated compressed features to the conversion operator of the graph neural network model includes the following steps:

[0031] Get the input nodes of the current level, apply the clustering algorithm to compress the input node features, and obtain the compressed features;

[0032] Repeat the previous step to compress the node features of other levels to obtain the corresponding compressed features;

[0033] A multi-level data reuse mechanism is adopted to apply transformation operators to the compressed features at each level.

[0034] Furthermore, the step of obtaining the input nodes of the current level and applying a clustering algorithm to compress the input node features to obtain compressed features includes:

[0035] Apply a clustering algorithm to the input nodes of the current level to obtain the node index and a set of representative features of each original node;

[0036] Input the representative features into the original transformation operation to obtain the representative output;

[0037] According to the node index representing the output and the original node, the node output is restored as the compressed feature.

[0038] In a second aspect, the present invention provides a data processing system based on graph neural network compression, comprising:

[0039] The data acquisition module is used to obtain historical data sets and preprocess them to obtain graph structure data sets;

[0040] The model training module is used to accelerate the training of the pre-built graph neural network model based on the graph structure dataset using a hybrid training method to obtain a trained graph neural network model;

[0041] The data processing module is used to process the target graph structure using the trained graph neural network model to obtain data processing results.

[0042] In a third aspect, the present invention provides a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any method.

[0043] In a fourth aspect, the present invention provides a computing device comprising: one or more processors and a memory, wherein the memory stores one or more programs and is configured to be executed by the one or more processors, and the one or more programs include instructions for executing any method.

[0044] The present invention has the following advantages due to the adoption of the above technical solution:

[0045] 1. This invention uses a hybrid training method to train graph neural networks. It accelerates the training process based on layered compression, enabling rapid convergence. It then fine-tunes the parameters of the graph neural network model using a conventional training model, achieving faster convergence speed and accuracy. This approach is suitable for applications requiring high accuracy.

[0046] 2. The present invention compresses the propagation operator and conversion operator of the graph neural network separately. The independent compression application of the two operators enables the present invention to be applied to various types of graph neural network structures, with a wider scope of applicability.

[0047] Therefore, the present invention can be widely applied in the field of big data processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Various other advantages and benefits will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiment below. The accompanying drawings are for illustration purposes only and are not to be considered as limiting the present invention. Throughout the drawings, the same reference numerals are used to denote the same components. In the drawings:

[0049] Figure 1 It is the existing graph neural network processing flow;

[0050] Figure 2a and Figure 2b This is a schematic diagram of the redundant operations of propagation and transformation in existing graph neural networks;

[0051] Figure 3This is a flow chart of a data processing method based on graph neural network compression provided in an embodiment of the present invention;

[0052] Figure 4 This is a flow chart of a compression-based graph neural network acceleration method provided in an embodiment of the present invention;

[0053] Figure 5 is a compression conversion operator provided in an embodiment of the present invention;

[0054] Figure 6 This is an example of local sensitive hashing provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field are within the scope of protection of the present invention.

[0056] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0057] like Figure 1 The following is the execution flow of the existing graph neural network. The input of a typical graph neural network includes a graph A and the feature information X corresponding to each node on the graph A. (0) The training process of a graph neural network consists of two operators: propagation and transformation. In the propagation phase, each node on the graph propagates its features to its neighbors. In the transformation phase, the node features undergo a nonlinear transformation through a deep neural network (DNN) layer. A graph neural network model is composed of a combination of propagation and transformation operators. The order of propagation and transformation varies slightly for different graph neural network model structures.

[0058] like Figure 2a and Figure 2b As shown, the present invention found that in the execution flow of existing graph neural networks, there is redundancy in both the propagation and conversion processes, including:

[0059] 1) Data redundancy during propagation

[0060] like Figure 2aAs shown in the figure, the core of the propagation operation is the aggregation of information between nodes through edges. Data redundancy primarily stems from repeated propagation to common neighbors in the graph structure. Specifically, when multiple nodes transmit information to the same set of common neighbor nodes, duplicate message paths are generated. For example, if nodes 1 and 2 are connected to nodes 3, 4, and 5, traditional propagation requires six passes: 1→3, 1→4, 1→5, 2→3, 2→4, and 2→5. However, the features of 1 and 2 can be first aggregated to a virtual node, which is then passed to nodes 3, 4, and 5, respectively, requiring only five passes (reducing the number of operations by 16.7%).

[0061] 2) Computational redundancy during conversion

[0062] like Figure 2b As shown, the transformation operation transforms node features through a neural network (such as the linear layer + activation function of GCN). Its computational redundancy mainly comes from the repeated calculation of similar features. Specifically, it manifests as: computational duplication caused by feature similarity: when multiple nodes have the same or highly similar features, they are calculated exactly the same way through the same transformation layer. For example, the input feature of nodes 1, 2, and 3 is x. After the linear transformation W and the activation function, the output features are exactly the same. In this case, only one transformation needs to be calculated, and the remaining nodes can directly reuse the results, eliminating 2 / 3 of the computation.

[0063] Based on this, some embodiments of the present invention provide a data processing method based on graph neural network compression. This method utilizes a graph neural network model to process data, and during the training process of the graph neural network model, the compression method is first applied to generate a compressed graph structure and compressed features; the compressed graph structure and compressed features are then applied to a propagation operator and a conversion operator; and finally, these two operators are arranged to form the desired graph neural network model. The present invention accelerates the training process of the graph neural network based on the compression method, reducing training time and training costs. The trained graph neural network model can be applied to multiple technical fields, including social networks, biomedical molecular modeling, etc., and in actual application processes, it can greatly improve data processing efficiency.

[0064] Correspondingly, in other embodiments of the present invention, a data processing system, device and medium based on graph neural network compression are provided.

[0065] Example 1

[0066] like Figure 3 and Figure 4 As shown, the present invention provides a data processing method based on graph neural network compression, which includes the following steps:

[0067] 1) Obtain historical data sets and preprocess them to obtain graph structure data sets;

[0068] 2) Based on the graph structure dataset, a hybrid training method is used to accelerate the training of the pre-built graph neural network model to obtain a trained graph neural network model;

[0069] 3) Use the trained graph neural network model to process the target graph structure and obtain the data processing results.

[0070] Furthermore, in step 1) above, the historical dataset of this embodiment may be data from social networks, biomedical molecular modeling, or the like, including a number of users and their relationships, with each user also including attribute information such as age, gender, and other attribute information. After preprocessing the collected data, a corresponding graph structure is obtained, including nodes and node feature information.

[0071] Furthermore, in step 2) above, this embodiment uses a hybrid training method to train the pre-built graph neural network model. First, a layered compression-based method is used to accelerate the training of the graph neural network model to achieve rapid convergence. Then, the model is switched to an uncompressed conventional training mode to fine-tune the parameters of the graph neural network model. This hybrid training method combines faster convergence speed with the ability to achieve optimal accuracy, ensuring that it is suitable for applications that require extremely high accuracy.

[0072] Specifically, the method includes the following steps:

[0073] 2.1) Compressing a given graph structure and applying the resulting compressed graph structure to the propagation operator of the graph neural network model to obtain a compression-based propagation operator;

[0074] 2.2) Compressing the node features and applying the generated compressed features to the transformation operator of the graph neural network model to obtain a compression-based transformation operator;

[0075] 2.3) Arrange the obtained compression-based propagation operators and transformation operators according to actual needs to obtain a preliminarily trained graph neural network model;

[0076] 2.4) Using the conventional training mode, fine-tune the graph neural network model that has completed the initial training to obtain a trained graph neural network model that meets the preset requirements.

[0077] Furthermore, in step 2.1) above, the present invention draws on the compression method of CompressGraph for graph compression. In the CompressGraph method, shared neighbors in the graph are recursively identified as virtual nodes, and then a graph containing original vertices and virtual vertices is generated. The present invention adopts a virtual node extraction method inspired by CompressGraph, but it is not feasible to directly apply it to GNN. In CompressGraph, vertices have no features, while in GNN, feature calculation is the main time cost, which requires efficient GPU operators for both node features and topology. The graph traversal in CompressGraph involves multiple iterations and uses a state array to record node status, while the number of nodes processed by GNN in each iteration is fixed. In addition, in the GNN model dominated by transformation operators, the propagation optimization of CompressGraph has limitations. Therefore, the present invention performs data and computation compression on propagation and transformation respectively, and develops a multi-step propagation engine to achieve efficient compressed graph propagation.

[0078] In this embodiment, the compression-based propagation operator includes two steps: offline and online. Since the graph structure can be reused after compression, after a given graph structure is given, the graph structure is compressed in the offline stage, and in the online stage, the compression-based propagation operator is applied to the compressed graph structure.

[0079] Specifically, the method includes the following steps:

[0080] 2.1.1) Compress the given graph structure using the CompressGraph method to generate a preliminary compressed graph containing original nodes and virtual nodes;

[0081] 2.1.2) Using a double-layer filtering method to filter virtual vertices and edges with low performance contribution in the initial compression graph, a filtered compression graph is obtained;

[0082] 2.1.3) Using a multi-step propagation engine, we apply the propagation operator of the graph neural network model on the filtered compressed graph to obtain a compression-based propagation operator.

[0083] Furthermore, in the above step 2.1.2), since the multi-step propagation engine is affected by the number of parallel rounds and the number of virtual nodes, in order to further improve the propagation efficiency, this embodiment filters the virtual nodes in the preliminary compression graph based on two dimensions. First, the concept of virtual node contribution is introduced to measure its impact on performance. The higher the contribution, the more beneficial it is. The contribution value of a virtual node is the product of the number of its input neighbors and the number of its output neighbors. For example, for a virtual node with 2 input neighbors and 3 output neighbors, its contribution value is 6. Second, the maximum depth of the compression graph is considered, which corresponds to the total number of parallel rounds.

[0084] Specifically, the method includes the following steps:

[0085] 2.1.2.1) Perform contribution-based filtering on the virtual nodes in the preliminary compressed graph G according to a preset contribution threshold until there are no virtual nodes in the compressed graph G whose contribution value is less than the contribution threshold.

[0086] Specifically, the method includes the following steps:

[0087] a) For each virtual node v in the preliminary compressed graph G, calculate its contribution value and insert it into an ascending priority queue;

[0088] b) Extract the virtual node with the smallest contribution value from the queue and compare its contribution value with the preset contribution threshold. If its contribution value is less than the contribution threshold, delete the virtual node and the corresponding edge from the compressed graph G, and update the contribution values ​​of its inbound and outbound neighbors in the queue;

[0089] c) Repeat step b) until there is no virtual node in the queue whose contribution value is less than the contribution threshold, and the contribution-based filtering process is terminated.

[0090] 2.1.2.2) Starting from the maximum depth of the compressed graph G, delete all virtual nodes layer by layer until the maximum depth of the compressed graph G reaches a preset depth threshold.

[0091] Furthermore, in step 2.1.3), to ensure that the propagation results on the compressed graph are consistent with those on the original graph, this embodiment uses a multi-step propagation engine to implement the propagation operator on the compressed graph. The specific process is as follows:

[0092] 2.1.3.1) Determine the depth of each node and edge on the filtered compressed graph according to preset rules.

[0093] In this embodiment, the depth of each node and edge on the filtered compressed graph is recorded as D, which is defined as:

[0094] a) For the original node on the graph, its depth is recorded as 1;

[0095] b) For a virtual node, its depth is the maximum depth of all its neighbors plus 1;

[0096] c) For all edges, their depth is the depth of the starting node.

[0097] 2.1.3.2) Based on the depth of each node and edge on the compressed graph, divide the nodes and edges with the same depth into the same subgraph. The number of all subgraphs is N.

[0098] 2.1.3.3) Based on the depth size, apply the propagation operator to propagate the features to the subgraphs with depths of 1 to N in sequence.

[0099] In this embodiment, the propagation operator is consistent with the original propagation operator. Their inputs are both a graph structure and a set of features. When propagating features, a space is first created to temporarily store the temporary features of virtual vertices. If the result of the propagation is the original vertex, it is placed in the space of the original vertex. If it is a virtual vertex, it is placed in the temporarily created space.

[0100] Furthermore, in step 2.2), the node features are compressed, and the generated compressed features are applied to the conversion operator of the graph neural network model, including the following steps:

[0101] 2.2.1) Obtain the input nodes of the current level and apply the clustering algorithm to compress the input node features to obtain compressed features;

[0102] 2.2.2) Repeat step 2.2.1) to compress node features at other levels to obtain corresponding compressed features;

[0103] 2.2.3) Adopt a multi-level data reuse mechanism and apply transformation operators to the compressed features at each level.

[0104] Furthermore, in the above step 2.2.1), if Figure 5 As shown, the following steps are included:

[0105] 2.2.1.1) Apply a clustering algorithm to the input nodes of the current level to obtain the node index and a set of representative features for each original node;

[0106] 2.2.1.2) Input the representative features into the original transformation operation to obtain the representative output;

[0107] 2.2.1.3) Restore the node output as the compressed feature based on the node index representing the output and the original node.

[0108] Furthermore, in step 2.2.1.1) above, the clustering algorithm of this embodiment adopts a projection-based locality sensitive hashing algorithm, which includes the following steps:

[0109] a) Randomly generate H vectors with a length equal to the width F of the vertex feature, and transpose them to obtain a random matrix R of size F*H;

[0110] b) Multiply the original node features with the random matrix R to obtain the output matrix;

[0111] In this embodiment, the original node features are expressed as N*F, where N is the number of nodes. After multiplying the original vertex features with the random matrix R, the output matrix obtained has a size of N*H;

[0112] c) Divide all elements in the output matrix, mark values ​​greater than 0 as 1, and values ​​less than or equal to zero as 0, to obtain a matrix P with all values ​​0 or 1, and the size is still N*H;

[0113] d) Combine each row of this 0 / 1 matrix into a 32-bit / 64-bit integer of size N*1;

[0114] e) For each row, if its numerical value is equal to that of another row, they are considered to be in the same category;

[0115] f) Complete clustering.

[0116] like Figure 6 Figure 1 shows an example of locality-sensitive hashing. In the figure, x1, x2, and x3 are three input features of dimension 4, and h1 and h2 are two random vectors. CompressGNN first projects x1, x2, and x3 onto the random vectors and then converts the projections into bit vectors. Since x1 and x2 have the same bit vectors, they are clustered into one class, while x3 belongs to a different class.

[0117] Furthermore, in step 2.2.3) above, the multi-level data reuse mechanism adopted in this embodiment includes:

[0118] The first level of reuse is in consecutive transformation operations. In this case, instead of clustering each transformation operation, the node indices obtained from the first clustering are reused in these transformations.

[0119] Second-level reuse involves reusing data across multiple rounds. During training, clustering isn't performed every round, but rather after several rounds. The frequency of second-level reuse is determined in real time based on the convergence rate. For example, if the convergence rate threshold is x, then re-clustering occurs if the convergence rate exceeds x; if it's less than x, the original features are retained.

[0120] Dynamically adjust the update frequency during training. Specifically, CompressGNN monitors the loss change between training rounds. As the loss change gradually converges, the update frequency is reduced accordingly.

[0121] The third level of reuse involves reusing node indexes. This method uses a similar approach to the second level of reuse to control whether to update the index. Similarly, the convergence rate can be used to infer changes in hidden layer features during training, dynamically determining the frequency of the second level of reuse.

[0122] Example 2

[0123] The above-mentioned embodiment 1 provides a data processing method based on graph neural network compression. Correspondingly, this embodiment provides a data processing system based on graph neural network compression. The system provided in this embodiment can implement the data processing method based on graph neural network compression in embodiment 1. The system can be implemented by software, hardware, or a combination of software and hardware. For example, the system may include integrated or separate functional modules or functional units to perform the corresponding steps in each method of embodiment 1. Since the system of this embodiment is basically similar to the method embodiment, the description process of this embodiment is relatively simple. For relevant points, please refer to the partial description of embodiment 1. The embodiment of the system provided in this embodiment is only illustrative.

[0124] The data processing system based on graph neural network compression provided in this embodiment includes:

[0125] The data acquisition module is used to obtain historical data sets and preprocess them to obtain graph structure data sets;

[0126] The model training module is used to accelerate the training of the pre-built graph neural network model based on the graph structure dataset using a hybrid training method to obtain a trained graph neural network model;

[0127] The data processing module is used to process the target graph structure using the trained graph neural network model to obtain data processing results.

[0128] Example 3

[0129] This embodiment provides a processing device corresponding to the data processing method based on graph neural network compression provided in this embodiment 1. The processing device can be a processing device for a client, such as a mobile phone, a laptop computer, a tablet computer, a desktop computer, etc., to execute the method of embodiment 1.

[0130] The processing device includes a processor, a memory, a communication interface, and a bus. The processor, memory, and communication interface are connected via the bus to facilitate communication between them. The memory stores a computer program executable on the processor. When the processor executes the computer program, it executes the data processing method based on graph neural network compression provided in Example 1.

[0131] Preferably, the memory may be a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk memory.

[0132] Preferably, the processor may be a central processing unit (CPU), a digital signal processor (DSP), or other general-purpose processors of various types, which are not limited here.

[0133] Example 4

[0134] The data processing method based on graph neural network compression of this embodiment 1 can be specifically implemented as a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing the data processing method based on graph neural network compression described in this embodiment 1.

[0135] Computer readable storage media can be tangible devices that hold and store instructions used by instruction execution devices. Computer readable storage media can be, for example, but not limited to, electronic storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any combination thereof.

[0136] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0137] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0138] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0139] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0140] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. A data processing method based on graph neural network compression, characterized in that: The following steps are involved: Obtain historical data sets and preprocess them to obtain graph structure data sets; Based on the graph structure dataset, a hybrid training method is used to accelerate the training of the pre-built graph neural network model to obtain a trained graph neural network model; The trained graph neural network model is used to process the target graph structure to obtain the data processing results.

2. The data processing method based on graph neural network compression according to claim 1, characterized in that: The hybrid training method is used to accelerate the training of the pre-built graph neural network model based on the graph structure dataset to obtain a trained graph neural network model, including: Compress the given graph structure and apply the generated compressed graph structure to the propagation operator of the graph neural network model to obtain a compression-based propagation operator; Compress the node features and apply the generated compressed features to the transformation operator of the graph neural network model to obtain a compression-based transformation operator; Arrange the obtained compression-based propagation operators and conversion operators according to actual needs to obtain a graph neural network model that has been preliminarily trained. The conventional training mode is used to fine-tune the graph neural network model that has completed the preliminary training to obtain a trained graph neural network model.

3. The data processing method based on graph neural network compression according to claim 2, characterized in that: The given graph structure is compressed, and the generated compressed graph structure is applied to the propagation operator of the graph neural network model to obtain the compression-based propagation operator, including: The given graph structure is compressed using the CompressGraph method to generate a preliminary compressed graph, which includes original nodes and virtual nodes. A double-layer filtering method is used to filter virtual vertices and edges with low performance contribution in the preliminary compression graph to obtain a filtered compression graph; A multi-step propagation engine is used to apply the propagation operator of the graph neural network model on the filtered compressed graph to obtain the compression-based propagation operator.

4. The data processing method based on graph neural network compression according to claim 3, characterized in that: The double-layer filtering method is used to filter virtual vertices and edges with low performance contribution in the preliminary compression graph to obtain a filtered compression graph, including: Perform contribution-based filtering on the virtual nodes in the preliminary compression graph according to a preset contribution threshold until there are no virtual nodes in the compression graph whose contribution value is less than the contribution threshold; Starting from the maximum depth of the compressed graph, all virtual nodes are deleted layer by layer until the maximum depth of the compressed graph reaches a preset depth threshold.

5. The data processing method based on graph neural network compression according to claim 3, characterized in that: The multi-step propagation engine is used to apply the propagation operator of the graph neural network model on the filtered compressed graph to obtain the compression-based propagation operator, including: Determine the depth of each node and edge on the filtered compressed graph according to preset rules; According to the depth of each node and edge on the compressed graph, nodes and edges with the same depth are divided into the same subgraph, and the number of all subgraphs is recorded as N; Based on the depth size, the propagation operator is applied to the subgraphs with a depth of 1 to N in sequence to propagate the features.

6. The data processing method based on graph neural network compression according to claim 2, characterized in that: The method of compressing node features and applying the generated compressed features to the conversion operator of the graph neural network model includes the following steps: Get the input nodes of the current level, apply the clustering algorithm to compress the input node features, and obtain the compressed features; Repeat the previous step to compress the node features of other levels to obtain the corresponding compressed features; A multi-level data reuse mechanism is adopted to apply transformation operators to the compressed features at each level.

7. The data processing method based on graph neural network compression according to claim 3, characterized in that: The step of obtaining the input nodes of the current level and applying a clustering algorithm to compress the input node features to obtain compressed features includes: Apply a clustering algorithm to the input nodes of the current level to obtain the node index and a set of representative features of each original node; Input the representative features into the original transformation operation to obtain the representative output; According to the node index representing the output and the original node, the node output is restored as the compressed feature.

8. A data processing system based on graph neural network compression, characterized in that: include: The data acquisition module is used to obtain historical data sets and preprocess them to obtain graph structure data sets; The model training module is used to accelerate the training of the pre-built graph neural network model based on the graph structure dataset using a hybrid training method to obtain a trained graph neural network model; The data processing module is used to process the target graph structure using the trained graph neural network model to obtain data processing results.

9. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any one of the methods of claims 1 to 7 .

10. A computing device, characterized in that include: One or more processors and a memory, wherein the memory stores one or more programs and is configured to be executed by the one or more processors, wherein the one or more programs include instructions for executing any one of the methods according to claims 1 to 7.

Images