A design space construction method for heterogeneous graph neural networks with multiple design dimensions

Abstract

The present invention discloses a method for constructing a design space of heterogeneous graph neural networks with multiple design dimensions, proposes a unified heterogeneous graph neural network framework, and defines the design space of heterogeneous graph neural networks based on this unified framework. The present invention makes up for the shortcomings of previous work that only evaluates heterogeneous graph neural networks from the model level, and provides researchers with a module-level evaluation perspective, which can further analyze which design dimensions play an important role in the performance of the model. We also built a platform Space4HGNN for the design space of heterogeneous graph neural networks, and based on this platform, we extracted the principled guidance of design models, provided standard evaluation and modular heterogeneous graph neural network implementation. The platform provides researchers with module-level evaluation and can provide users with a simple interface to help them evaluate the impact of different design dimensions.

Description

Technical Field

[0001] The present invention relates to the field of graph processing technology, and in particular to a method for constructing a design space of a heterogeneous graph neural network with multiple design dimensions. Background Art

[0002] Graph neural networks (GNNs) aim to apply neural networks to graph-structured data. Unlike homogeneous graph neural networks, heterogeneous graph neural networks (HGNNs) need to handle structural heterogeneity and capture the rich semantics of heterogeneous graphs. Based on their strategies for handling heterogeneity, heterogeneous graph neural networks can be roughly divided into two categories: heterogeneous graph neural networks based on one-hop neighbor aggregation (similar to traditional homogeneous graph neural networks) and heterogeneous graph neural networks based on meta-path neighbor aggregation (used to mine semantic information).

[0003] 1) Heterogeneous Graph Neural Network Based on One-Hop Neighbor Aggregation

[0004] To handle heterogeneity, this class of heterogeneous graph neural networks typically incorporates a specific type of convolution. Similar to homogeneous graph neural networks, aggregation occurs within one-hop neighbors. RGCN (Relational Graph Convolutional Network), the earliest work and an extension of GCN (Graph Convolutional Network), assigns different weight matrices to different relation types and aggregates one-hop neighbors. The emergence of many homogeneous graph neural network variants has inspired the emergence of more heterogeneous graph neural networks. HGConv (Heterogeneous Graph Convolution) then implements dual aggregation of one-hop neighbors based on GATConv (Graph Attention Network Convolution). A more recent work, SimpleHGN (Simple Heterogeneous Graph Network), designs relationship-specific weight matrices and embeddings to represent heterogeneous attention on each edge. In addition, some early models, such as HGAT (Heterogeneous Graph Attention Network), HetSANN (Heterogeneous Graph Structural Attention Neural Network), and HGT (Heterogeneous Graph Transformer), generate heterogeneous variant models of GAT by assigning heterogeneous attention to nodes or edges.

[0005] 2) Heterogeneous Graph Neural Network Based on Meta-path Neighbor Aggregation

[0006] Another type of heterogeneous graph neural network captures high-order semantic information by manually defining meta-paths. Unlike the former, the aggregation process occurs in neighbors connected by meta-paths. As a pioneering work, HAN first used node-level attention to aggregate nodes connected by the same meta-path, and used semantic-level attention to fuse information from different meta-paths. Since the meta-path subgraph ignores all intermediate nodes, MAGNN (Metapath Aggregated Graph Neural Network) proposes to aggregate all nodes in the meta-path instance to ensure that information is not missed. Although meta-paths contain rich semantic information, the selection of meta-paths requires prior knowledge and determines the performance of heterogeneous graph neural networks. GTN (Heterogeneous Graph Transformer Network) constructs new graphs by automatically learning meta-paths.

[0007] At present, heterogeneous graph neural networks have been successfully applied to various tasks, but due to the diversity of architectures and application scenarios, we cannot accurately understand the importance of different design dimensions of heterogeneous graph neural networks. In addition, it is increasingly difficult for researchers in this field to compare existing methods and propose new methods. The main reason is that previous evaluations were mostly conducted from a model-level perspective, and due to the differences in architecture design and application scenarios, we cannot accurately understand the importance of each component. For example, the difference between existing relationship-based models (such as RGCN) and meta-path-based models (such as HAN) lies not only in the receptive field, but also in the choice of graph convolution. Under model-level evaluation, it will be unknown which design dimension will produce better performance.

[0008] GraphGym systematically studies a general neural network design space across a wide range of tasks and proposes three core components: the graph neural network design space, the graph neural network task space, and the design space evaluation. (1) The graph neural network design space contains important architectural design aspects that researchers typically consider when building models. (2) The graph neural network task space and a task similarity metric provide guidance for migrating architectural designs across similar tasks. (3) A design space evaluation method can extract guiding principles from a wide range of model and task combinations. GraphGym also provides a platform for exploring different neural network designs and tasks, including modular implementations of graph neural networks, standard graph neural network evaluations, and reproducible large-scale experiment management.

[0009] However, much existing work measures progress in this field by evaluating models. This approach assesses the quality of models from a model-level perspective, failing to accurately capture which components within the model play a key role and significantly impact the final performance. Therefore, for heterogeneous graph neural networks, no research has yet conducted in-depth analysis of the architectural design of heterogeneous graphs, nor has it addressed the impact of heterogeneous graph neural network design on performance from a module-level perspective, providing little guidance for designing new models. Summary of the Invention

[0010] In response to the above technical problems, the present invention aims to extensively explore the design space of heterogeneous graph neural networks involving many design dimensions, evaluate different design architectures, and provide a design space construction method for heterogeneous graph neural networks with multiple design dimensions, so as to realize evaluation from a module-level perspective and analyze the effects of different architectures or design principles on the model.

[0011] In order to achieve the above object, the present invention provides the following technical solutions:

[0012] A method for constructing a design space of a heterogeneous graph neural network with multiple design dimensions includes the following steps:

[0013] S1. We propose a unified heterogeneous graph neural network framework consisting of three major components: heterogeneous linear transformation, heterogeneous graph transformation, and heterogeneous message passing layer.

[0014] S2. Based on a unified heterogeneous graph neural network framework, three heterogeneous graph neural network model families are proposed. Three corresponding subspaces are designed for these three model families, namely the homogeneous model family, the relational model family, and the meta-path model family. The design space of heterogeneous graph neural networks is defined, which contains two design dimensions: the design dimension shared with homogeneous graph neural networks and the design dimension unique to heterogeneous graph neural networks.

[0015] S3. Space4HGNN, a design space building platform for heterogeneous graph neural networks, is used for modular heterogeneous graph neural network implementation and standardized heterogeneous graph neural network evaluation.

[0016] Furthermore, the heterogeneous linear transformation described in step S1 maps the features or representations of the heterogeneous nodes into a shared feature space. The linear transformation is expressed as follows:

[0017]

[0018] Among them, h i and h′ i They are node v i The original features and mapped features, f V (v i ) is the node v iThe node type.

[0019] Furthermore, the heterogeneous graph transformation described in step S1 provides four conversion methods for heterogeneous graph data before aggregation, namely: relationship subgraph extraction, metapath subgraph extraction, construction of a hybrid subgraph of two graphs, and homogenization of heterogeneous graphs.

[0020] Furthermore, for relation subgraph extraction, the relation type and the original heterogeneous graph are taken as input.

[0021] Furthermore, for the extraction of meta-path subgraph, the predefined meta-path and the original heterogeneous graph are taken as input.

[0022] Furthermore, the heterogeneous message passing layer described in step S1 defines two aggregation methods for heterogeneous graph neural networks, namely: direct aggregation and dual aggregation; the aggregation process of direct aggregation is to directly aggregate adjacent neighbors without distinguishing between node types; dual aggregation is an aggregation process applied to different subgraphs at the micro and macro levels. The micro-level aggregation is to aggregate node features within the same relationship type and generate type-specific features in the relationship or meta-path subgraph, while the macro-level aggregation is to aggregate type-specific features between different relationships.

[0023] Furthermore, the formula for direct aggregation is as follows:

[0024]

[0025] in Represents node v i The neighbor type of node j is the same as that of node j.

[0026] Furthermore, the design dimensions shared with the homogeneous graph neural network in step S2 involve 12 design dimensions, which are divided into three aspects: intra-layer, inter-layer and training settings; the intra-layer design dimensions are the same as those of the homogeneous graph neural network. The heterogeneous graph neural network contains several heterogeneous neural network layers, and each layer has a different design size; the inter-layer design dimensions involve message passing, pre-processing and post-processing layers; the training setting design dimensions involve optimizers, learning rates and number of training rounds.

[0027] Furthermore, the design dimensions unique to the heterogeneous graph neural network in step S2 involve three design dimensions: the qualitative model family uses direct aggregation combined with any graph convolution; the relational model family uses relational subgraph extraction and dual aggregation; and the meta-path model family uses meta-path subgraph extraction and dual aggregation.

[0028] Furthermore, the Space4HGNN described in step S3 is implemented using the OpenHGNN toolkit using PyTorch and DGL.

[0029] Compared with the prior art, the present invention has the following beneficial effects:

[0030] This invention provides a method for constructing a design space for heterogeneous graph neural networks with multiple design dimensions. It proposes a unified heterogeneous graph neural network framework consisting of three major components: heterogeneous linear transformations, heterogeneous graph transformations, and a heterogeneous message passing layer. Based on this unified framework, the design space for heterogeneous graph neural networks is defined. This design space primarily encompasses two aspects: design dimensions shared with homogeneous graph neural networks and design dimensions unique to heterogeneous graph neural networks. This invention addresses the shortcomings of previous work that evaluated heterogeneous graph neural networks solely at the model level, providing researchers with a module-level evaluation perspective that further analyzes which design dimensions play a significant role in model performance. We also constructed a platform, Space4HGNN, for the design space of heterogeneous graph neural networks. Based on this platform, we have refined principled guidance for design models, provided standardized evaluation, and provided modular heterogeneous graph neural network implementations. This platform offers researchers module-level evaluation and provides users with a simple interface to help them evaluate the impact of different design dimensions, such as high-level architectural design and design principles, enabling extensive research on these design dimensions. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments described in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.

[0032] Figure 1 Schematic diagram of heterogeneous linear transformation provided by an embodiment of the present invention.

[0033] Figure 2 A schematic diagram of heterogeneous graph transformation provided in an embodiment of the present invention, wherein the left figure is an original graph composed of four adjacency matrices, and the four figures on the right are schematic diagrams of four methods of heterogeneous graph transformation.

[0034] Figure 3 A schematic diagram of the direct aggregation process of the heterogeneous message transmission layer provided by an embodiment of the present invention.

[0035] Figure 4 Schematic diagram of the dual aggregation process of the heterogeneous messaging layer provided in an embodiment of the present invention, where the thin solid lines with arrows represent the micro-level aggregation process applied to different subgraphs, and the thick solid lines with arrows represent the macro-level aggregation process.

[0036] Figure 5 The solid-line box represents the unique design dimensions of the heterogeneous graph neural network, and the dashed-line box represents the design dimensions shared with GraphGym. DETAILED DESCRIPTION

[0037] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0038] First, we introduce some definitions of graph neural networks and heterogeneous graphs.

[0039] Definition 1 (Message Passing GNN). Message passing GNN learns a representation vector for each node v after being transformed by the Lth message passing layer. d L represents the output dimension of the Lth message passing layer. The message passing paradigm defines the following node-level and edge-level computations for each layer:

[0040] Edge level:

[0041] in, Represents node v i Neighbor nodes, φ is a message function defined on each edge, which generates a message by aggregating the features of its incident nodes. ij Represents the slave node v j to v i An edge of :

[0042] Node level:

[0043] in Represents node v i neighbor nodes, ψ is an update function defined on each node, and the aggregation function ρ is used to aggregate its incoming messages to update the node representation.

[0044] Example: GraphSAGE can be formalized as a message-passing GNN, where the message function is The update function is

[0045] Definition 2 (Heterogeneous Graph). A heterogeneous graph is represented as Contains a node set and an edge set ε. The heterogeneous graph also has a mapping function with node types and a mapping function for an edge type About, among which and Represents a set of node types and edge types. Each node There is a node type f v (v i )∈T v , similarly, each edge e connecting node i and node j ij ∈ε satisfies when or , it is a heterogeneous graph, otherwise it is a homogeneous graph.

[0046] Definition 3 (Relationship subgraph). A heterogeneous graph can also be represented by a set of adjacency matrices To indicate that Represents the number of edge types. Is an adjacency matrix, where when there is a k-th type, from node v i To node v j When the edge of k [i, j] is non-zero. n s and n t is the number of source nodes and target nodes corresponding to edge type k. The relationship subgraph of the kth edge type is the adjacency matrix A k 's subgraph.

[0047] Definition 4 (Metapath). Metapath Defined as Describe two nodes v1 and v in the form of l+1 The compound relationship between The path of The meta-path representing the first type of relationship, Represents the relational composition operator.

[0048] Definition 5 (Meta-path subgraph). Given a meta-path According to the relationship The adjacency matrix can be obtained by multiplication of the adjacency matrix

[0049]

[0050] The concept of a meta-path includes multi-hop connections, and a meta-path subgraph is the matrix multiplication of multiple relation subgraphs. Therefore, a relation subgraph is a special case of a meta-path subgraph, consisting of only one relation subgraph. When the start and end node types of a meta-path are the same, the meta-path subgraph is a homogeneous graph; otherwise, it is a bipartite graph.

[0051] The design space construction method of a heterogeneous graph neural network with multiple design dimensions of the present invention comprises the following steps:

[0052] 1. Unified Heterogeneous Graph Neural Network Framework

[0053] By analyzing the aggregation process of underlying graph data and existing heterogeneous graph neural networks, we propose a unified heterogeneous graph neural network framework, which consists of the following three main parts:

[0054] Heterogeneous linear transformation: Mapping heterogeneous features or representations into a shared feature space.

[0055] Heterogeneous Graph Transformation: Four transformation methods are provided for heterogeneous graph data before aggregation.

[0056] Heterogeneous Message Passing Layer: We define two aggregation methods that are suitable for most heterogeneous graph neural networks.

[0057] 1. Heterogeneous linear transformation

[0058] Due to the heterogeneity of nodes, different types of nodes have different semantic features and even different dimensional sizes. Therefore, for each type of node (for example, node v i The node type is f v (v i ), we design a specific type of linear transformation to map the features (or representations) of different types of nodes into the same feature space. The linear transformation is expressed as follows:

[0059]

[0060] where h i and h′ i They are node v i The original features and mapped features of .

[0061] like Figure 1 As shown in Figure 2, for nodes with features, we use a specific type of linear transformation to map node features. For nodes without features or full of noise, they can be assigned embeddings as trainable vectors, which is equivalent to assigning them a one-hot encoding vector combined with a linear transformation.

[0062] 2. Heterogeneous Graph Transformation

[0063] One-hop neighbor-based aggregation typically applies graph convolutional layers to the original graph, but meta-path neighbor-based aggregation typically operates on a constructed meta-path subgraph. As described in Definition 5, a relation subgraph is a special case of a meta-path subgraph (note that the original graph is a special case of a relation subgraph). To unify the two, we propose a component to abstract the receptive field selection process, which determines which type of nodes (one-hop neighbors or meta-path neighbors) are aggregated.

[0064] like Figure 2 As shown, we designate a separate phase for graph construction, called heterogeneous graph transformation, and classify it into the following four methods: (i) relation subgraph extraction, (ii) meta-path subgraph extraction, (iii) hybrid subgraph extraction that builds two graphs, and (iv) homogenization of heterogeneous graphs (but still preserving the node and edge type mapping). v and f e ).

[0065] For relation or meta-path subgraph extraction, we only need to input some relation types or predefined meta-paths and the original heterogeneous graph. The heterogeneous graph transformation decouples the receptive field selection process from the message passing process described below.

[0066] 3. Heterogeneous messaging layer

[0067] Traditional heterogeneous graph neural network classification methods do not find enough commonality from an implementation perspective, leading to difficulties in defining the design space and finding new models. Therefore, we instead propose a classification method based on model aggregation.

[0068] 1) Direct Aggregation: The line with the normalized attention coefficient represents the aggregation process. This aggregation process directly aggregates adjacent neighbors and does not distinguish between node types.

[0069] The baseline models for heterogeneous graph neural network models are GCN, GAT, and other GNNs for homogeneous graphs. Simple homogeneous GNNs (such as GCN and GAT) have been underestimated due to improper settings. As shown in Table 1, the models in this series all use direct aggregation and a single graph convolution layer.

[0070] like Figure 3 As shown, below we will explain it under the GNN representation of message passing and take GAT as an example.

[0071] The message function is The feature of node i in the (L+1) layer is defined as:

[0072]

[0073] Among them, W is the trainable weight matrix, is node v i Neighbor, α ij is node v i and v j The standardized attention coefficient between is defined as:

[0074]

[0075] Node v i With its neighbors The relevance of ij express.

[0076] Change e ij Other heterogeneous variants of GAT can be obtained in the form of , which we summarize in Table 1.

[0077] Table 1 Direct aggregation of attention mechanism

[0078]

[0079] 2) Dual Aggregation: We define two parts of dual aggregation: micro-level (within-type) and macro-level (between-type) aggregation.

[0080] like Figure 4 As shown, micro-level aggregation aggregates node features within the same relation type to generate type-specific features in the relation / meta-path subgraph, while macro-level aggregation aggregates type-specific features across different relations. When multiple relations have the same target node type, their type-specific features are aggregated by macro-level aggregation.

[0081] In general, each relation / meta-path subgraph utilizes the same micro-level aggregation (e.g., graph convolution layers from GCN or GAT). In fact, we can apply different homogeneous graph convolution operations for different subgraphs in our framework. Compared with heterogeneous graph convolution in direct aggregation, multiple homogeneous graph convolutions combined with macro-level aggregation is another form of heterogeneous graph convolution. There is a small difference between heterogeneous graph convolution with direct aggregation and heterogeneous graph convolution with dual aggregation. We formulate and define direct aggregation as follows:

[0082]

[0083] in Represents node v i The neighbor type of node j is the same as that of node j.

[0084] Examples: HAN and HGConv. In HAN, node-level attention is equivalent to the micro-level aggregation of GATConv, while semantic-level attention is macro-level attention aggregation, which is the same as HGConv. HGConv uses relational subgraphs, that is, aggregating one-hop neighbors, but HAN extracts multiple meta-path subgraphs, that is, aggregating multi-hop neighbors. Based on the heterogeneous graph transformation described above, the constructed graph can be a mixture of meta-path subgraphs and relational subgraphs. Therefore, dual aggregation can also operate on mixed subgraphs to aggregate neighbors of different hops.

[0085] 2. Design Space of Heterogeneous Graph Neural Networks

[0086] like Figure 5As shown in Figure 2, we define the design space of heterogeneous graph neural networks and describe it from two perspectives: common designs in GraphGym and unique designs extracted from heterogeneous graph neural networks. Although GraphGym extracts relevant principles for designing well-performing graph neural networks, we still retain relevant design dimensions to clarify whether they are still effective in heterogeneous graph neural networks. However, the design dimensions used in graph neural networks are not sufficient to define the design space of heterogeneous graph neural networks, and there are also some unique dimensions in heterogeneous graph neural networks.

[0087] Based on the unified framework proposed above, we propose three families of heterogeneous graph neural network models that can reproduce most of the original heterogeneous graph neural network models and extend them into new variants by using different choices in unique design dimensions. To highlight their role in heterogeneous graph neural networks, we design three corresponding subspaces for these three model families: the homogeneous model family, the relational model family, and the meta-path model family. Furthermore, our design space is constructed as a platform to provide researchers with modular heterogeneous graph neural network implementations, which will be introduced at the end.

[0088] 1. Design dimensions shared with GraphGym

[0089] The design shared with GraphGym involves 12 design dimensions, which are divided into three aspects: within-layer, between-layer, and training settings. Table 2 shows these dimensions and the corresponding choices.

[0090] Table 2 Design dimensions shared with GraphGym

[0091]

[0092] 1) Intra-layer design: Similar to homogeneous graph neural networks, heterogeneous graph neural networks contain several heterogeneous neural network layers, each of which can have different design sizes.

[0093] like Figure 2 As shown in Figure 1, the heterogeneous graph neural network layer used has an aggregation layer, which involves unique design dimensions discussed later, followed by a series of modules: (1) batch normalization BN(·), (2) dropout method DROP(·), (3) nonlinear activation function ACT(·), and (4) L2 normalization L2-Norm(·).

[0094] Layers such as skip connections and post-processing are essential design dimensions. Heterogeneous graph neural network layers iteratively transform features, leading to problems such as vanishing gradients, overfitting, and oversmoothing. The latter problem is considered an obstacle to stacking deeper neural network layers. We investigated two skip connection options: SKIP-SUM and SKIP-CAT, using STACK as a baseline for comparison.

[0095] 3) Training Settings: As part of deep learning, we also need to analyze the design dimensions of the training settings, such as the optimizer, learning rate, and number of training epochs. In addition, this also includes the hidden dimensions involving trainable parameters.

[0096] 2. Unique Design of Heterogeneous Graph Neural Network

[0097] Through this unified framework, we attempt to transform modular components into distinct design dimensions within graph neural networks. A collection of neural network architectures, often sharing high-level architectural structures or design principles (e.g., residual connections), can be abstracted as a model family. Using this framework, we identify three model families within heterogeneous graph neural networks. As shown in Table 3, these three model families involve three design dimensions and their candidate choices.

[0098] Table 3 Unique design of heterogeneous graph neural network

[0099]

[0100] 1) Homogeneous Model Family: The homogeneous model family combines any graph convolution with direct aggregation. We use the term homogeneous here because all models in this family homogenize heterogeneous graphs and then apply direct aggregation. Homogeneous Graph Neural Networks: The heterogeneous variants of the GAT mentioned above all belong to this model family. In heterogeneous graph neural network papers, homogeneous heterogeneous graph neural networks are often evaluated as the basic baseline. Ignoring the heterogeneity of nodes and edges and feeding the original graph into a homogeneous graph neural network makes intuitive sense. Although this loses type information, simple homogeneous graph neural networks have been shown to outperform some existing heterogeneous graph neural networks, meaning they cannot be ignored and should be considered as a model family. We selected four graph convolutions from their respective original models for analysis: GraphConv, GATConv, SageConv-mean, and GINConv.

[0101] 2) Relational model family: This model family applies relational subgraph extraction and dual aggregation. The first heterogeneous graph neural network model RGCN is a typical example in the relational model family. Its dual aggregation includes SageConv-mean aggregation at the micro level and Sum aggregation at the macro level. HGConv is a combination of GATConv and Attention. We can derive other designs by enumerating the combinations of micro-level and macro-level aggregation. In our experiments, we set the micro-level aggregation to be the same as the graph convolution in the homogeneous model family, and the macro-level aggregation is selected from Mean, Max, Sum, and Attention.

[0102] 3) Metapath Model Family: This family of models applies metapath subgraph extraction and dual aggregation. The HAN model has the same dual aggregation as HGConv in the relational model family, but the subgraph extraction method is different. The candidates for micro- and macro-level aggregation are the same as those in the relational model family.

[0103] 3. Space4HGNN: A Design Space Platform for Heterogeneous Graph Neural Networks

[0104] We also developed a new platform, Space4HGNN, for exploring the design of heterogeneous graph neural networks. Space4HGNN is implemented using PyTorch and DGL using the OpenHGNN toolkit, and can significantly advance the research field of heterogeneous graph neural networks. Furthermore, it provides a standardized evaluation process for heterogeneous graph neural networks. The advantages of Space4HGNN are as follows:

[0105] 1) Modular heterogeneous graph neural network implementation: The design space is easily extensible, allowing future developers to insert more design dimension choices (for example, a new graph convolution layer or a new macro aggregation). In addition, it is easy to import new design dimensions into Space4HGNN, such as the scoring function in link prediction.

[0106] 2) Standardized heterogeneous graph neural network evaluation: Space4HGNN provides a standardized evaluation process for different architecture designs and heterogeneous graph neural network models, which allows us to easily evaluate different datasets in different tasks.

[0107] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A design space construction method for heterogeneous graph neural networks with multiple design dimensions, characterized by: The following steps are involved: S1. A unified heterogeneous graph neural network framework is proposed, which consists of three major components: heterogeneous linear transformation, heterogeneous graph transformation, and heterogeneous message passing layer. The heterogeneous linear transformation described in step S1 maps the features or representations of heterogeneous nodes into a shared feature space. The linear transformation is expressed as follows: Among them, h i and h i ′ They are node v i The original features and mapped features, f v (v i ) is the node v i The node type; The heterogeneous graph transformation described in step S1 provides four transformation methods for heterogeneous graph data before aggregation, namely: relation subgraph extraction, meta-path subgraph extraction, construction of a hybrid subgraph of the two graphs, and homogenization of heterogeneous graphs. For relation subgraph extraction, the relation type and the original heterogeneous graph are used as input; for meta-path subgraph extraction, the predefined meta-path and the original heterogeneous graph are used as input. The heterogeneous message passing layer described in step S1 defines two aggregation methods for heterogeneous graph neural networks: direct aggregation and dual aggregation. Direct aggregation is the aggregation process of directly aggregating adjacent neighbors without distinguishing between node types. Dual aggregation is an aggregation process applied to different subgraphs at the micro and macro levels. Micro-level aggregation aggregates node features within the same relationship type to generate type-specific features in the relationship or meta-path subgraph, while macro-level aggregation aggregates type-specific features between different relationships. S2. Based on a unified heterogeneous graph neural network framework, three heterogeneous graph neural network model families are proposed, and three corresponding subspaces are designed for these three model families, namely the homogeneous model family, the relational model family, and the meta-path model family. The design space of heterogeneous graph neural networks is defined, which contains two design dimensions: the design dimensions shared with homogeneous graph neural networks and the design dimensions unique to heterogeneous graph neural networks. The design dimensions shared with homogeneous graph neural networks involve 12 design dimensions, which are divided into three aspects: intra-layer, inter-layer, and training settings. The intra-layer design dimensions are the same as those of homogeneous graph neural networks. Heterogeneous graph neural networks contain multiple heterogeneous neural network layers, and each layer has a different design size. The inter-layer design dimensions involve message passing, pre-processing, and post-processing layers. The training setting design dimensions involve optimizers, learning rates, and the number of training rounds. The design dimensions unique to heterogeneous graph neural networks involve three design dimensions: the homogeneous model family uses direct aggregation combined with any graph convolution; the relational model family uses relational subgraph extraction and dual aggregation; and the metapath model family uses metapath subgraph extraction and dual aggregation. S3. Space4HGNN, a design space building platform for heterogeneous graph neural networks, is used for modular heterogeneous graph neural network implementation and standardized heterogeneous graph neural network evaluation.

2. The design space construction method of a heterogeneous graph neural network with multiple design dimensions according to claim 1 is characterized in that: The formula for direct aggregation is as follows: in Represents node v i The neighbor type of node j is the same as that of node j.

3. The design space construction method of a heterogeneous graph neural network with multiple design dimensions according to claim 1 is characterized in that: The Space4HGNN described in step S3 is implemented using the OpenHGNN toolkit with PyTorch and DGL.

Images