GB/T 42382.3-2026Information technology - Neural network representation and model compression - Part 3: Graph neural networks (English PDF)

6.1 Basic Definitions

6.1.1 Overview This document defines the data structures and various basic data types.

6.1.2 Data Structure Definition This document defines a message as the basic unit of a data structure, and fields as the basic elements that constitute a message, as shown in Figure 2. As shown. Figure

2 Data Structure Definition Specification Messages consist of a series of fields with specific labels (fieldlabel), types (fieldtype), names (fieldname), and numbers (field). The data consists of fields (e.g., numbers). The message itself can be used as the field type to enable nested data structures.

a) Field Label. Field labels are used to define the attributes and usage rules of a field. By default, all fields are... It is considered optional and does not require explicit declaration. This approach aims to simplify the use of programming languages and ensure backward compatibility. When a field needs to represent multiple values, the "repeated" keyword should be used to indicate that the field can contain one or more values. The values above.

b) Field Type. The field type defines the data type that the field can store. This type includes int32, float, etc. Standard data types such as double, bool, and string, as well as composite data types such as enumerations, maps, and messages. type.

c) Field Name. The field name is a unique name specified for the field in the message definition, used to identify and access the message. The specific details of the information.

d) Field number. The field number is a positive integer used to uniquely identify a field in the binary format of the message. This number is crucial in the message serialization and deserialization process, ensuring that even if field names differ across languages... Similarly, the structure of the data can still be correctly understood.

6.2.4 Definition of Dynamic Graphs The definition of a dynamic graph is shown in Table 14.

6.2.5 Definition of Coordinate Graph The definition of a coordinate graph is shown in Table 15.

6.3 Basic Graph Operations The basic operation definitions for graph operations are shown in Tables 16 to 49. The definition of the degree operation is shown in Table 16.

7.1.1 Definition of Computational Graph

7.1.1.1 Defining the computation graph class The definition classes for computational graphs are shown in Table 50. Version Definition

7.1.1.3 The version definition is shown in Table 52.

7.1.1.4 Function Library Definition The function definition library (FunctionDefLibrary) is shown in Table 53.

7.1.1.5 Calculation Graph Debugging Information The definition of computation graph debugging information (GraphDebugInfo) is shown in Table 58.

7.1.2 Definition of Graph Neural Network Model Structure The graph neural network model structure definition (Model) is shown in Table 61.

7.2 Basic Operators

7.2.1 Overview Basic operators are the fundamental components for passing and aggregating messages in graph neural networks, and are divided into message passing operators (MessagePassingOperator-) The three operators are. pooling operator, normalization operator, and pooling operator.

7.2.2 Message Passing Operator

7.2.2.1 Overview The message passing operator is a core operation in graph neural networks, simulating the propagation and interaction of messages within a graph structure. Its computation... The process includes three sub-operators. message operator, aggregate operator, and update operator. (operator). Message-passing-based graph neural network convolution operators can be combined using message operators, aggregation operators, and update operators. accomplish.

7.2.2.2 Message Operators Message operators are operations defined on each edge that generate information by combining the characteristics of the edge with the characteristics of its attached nodes. The definition of the send_message operation is shown in Table 62.

7.2.2.4 Update Operator Update operators are a key component in the message passing process. They are defined on each node and process aggregated messages or their own characteristics. To update node characteristics.

7.2.2.5 Convolution Operator This document defines various classic graph convolution operators, and the specific definitions are shown in Tables 69 to 121. The GCNConv operation definition is shown in Table 69.

7.5.4 Graph Clustering Graph clustering in graph neural networks refers to identifying graphs with similar topological structures, similar node attribute distributions, or other associations. The task of classifying each graph into a predefined category is based on the graph's properties and topology. Graph neural networks utilize node attributes and graph topology to... Learn the representation of the graph and use it to calculate the probability or score of belonging to different categories.

7.5.5 Graph Matching Graph matching in graph neural networks refers to learning representations of nodes and edges in a graph using graph neural networks, based on the representations of nodes and edges. The method considers both node similarity and edge similarity to establish meaningful structural correspondences between nodes in two or more graphs. The graph matching models are shown in Figures 210 to 211.

8.1 Image Data Compression

8.1.1 Overview Graph data compression refers to compressing graph data, which can reduce the overhead of graph data storage and transmission, and improve the efficiency of graph data processing. Scalability. Graph data compression can be categorized into graph topology compression and graph feature data compression. See Tables 214-217 for specific definitions.

8.1.2 Graph Topology Compression Graph topology compression refers to transforming the representation of a graph's topology into a more compact and efficient form. This involves compressing the graph's topology. Compression can reduce the storage space occupied by graph topology data and improve processing efficiency.

8.1.3 Feature Data Compression Graph feature data compression refers to compressing the feature data of nodes, edges, or the entire graph to reduce the space occupied by the feature data. This method improves the computational efficiency of graph neural networks in processing feature data. The node and edge features of graph data can contain multiple values, such as node... Feature vectors of points, feature vectors of edges, etc. Feature data compression uses appropriate compression methods and algorithms to compress the attributes of vectors or arrays. The representation is transformed into a more compact form to reduce data storage space.

8.1.4 Image Sampling

8.1.4.1 Overview Based on the different levels of graph sampling, graph sampling methods are divided into three categories.

a) Node sampling. Node sampling samples the local neighbors of a single node. Node sampling selects a fixed number of neighbors for each node in the graph. The neighbors of a node are used to compute its features. This type of method allows the model to reduce the computational overhead caused by factors beyond the local structure of each node. Considering the computational burden caused by all neighbors.

b) Hierarchical sampling. Hierarchical sampling samples the neighbors of multiple nodes at each layer of the GNN. Compared to node sampling, this method... This method can ensure sampling efficiency while avoiding the problem of exponential growth in the number of neighbors. The number of nodes in hierarchical sampling depends on the number of layers. This linear relationship reduces overall memory overhead and improves overall computation speed, effectively reducing computational overhead even under resource-constrained conditions. Deep GNN models are processed.

c) Subgraph Sampling. Subgraph sampling obtains a subgraph containing selected nodes and edges from the original graph, and then performs full sampling on this subgraph. Training of graph-based GNNs. This type of method can partition large-scale graphs into multiple smaller, more tractable subgraphs, reducing computation. Load, while maintaining the global properties of the graph.

8.1.4.2 Node Sampling Node sampling refers to sampling the neighbors of a target node to reduce the amount of data in the input graph, which can reduce memory consumption and accelerate the model. train. The implementation process and details of this method are as follows:

a) Input. Graph data including node features, adjacency matrix, edge features, etc., with several target nodes selected;

b) Use a sampling strategy to sample neighbors;

c) Output. Sampled graph data, the node number of each node in the original graph, and the node number of the target node in the output graph.

8.1.4.3 Hierarchical Sampling Hierarchical sampling refers to sampling a fixed number of neighboring nodes at each layer of a GNN. Hierarchical sampling can address the issue of sampling as network depth increases. To address the issue of exponential growth in the number of neighbors, this approach reduces memory consumption and accelerates model training.

8.1.5 Graph Data Compression Metrics Graph data compression metrics are indicators used to evaluate the effectiveness and performance of graph data compression algorithms. The effectiveness of graph data compression algorithms is primarily determined by... The performance of graph data compression algorithms should be reflected in the compressed space usage, data representation capabilities, and recoverability. The performance of graph data compression algorithms is mainly reflected in the compression calculation... The time overhead and scalability of this method are scalable. Commonly used metrics for graph data compression include graph edit distance, compression ratio, information entropy, and compression ratio. time.

8.2 Model Quantization and Pruning

8.2.1 Quantification Quantization of graphical models refers to converting the parameters in the model into fixed-point or floating-point numbers. Quantization can reduce the model's storage space and computation. Reduce complexity and improve model efficiency and inference speed.