Topology structure design method and system of graph neural network

Abstract

A method and system for designing a topology of a graph neural network are provided. Graph data is obtained. The graph data is input into a directed acyclic feature fusion framework, where the feature fusion framework includes an input block, a predetermined number of ordered aggregation blocks, and an output block, where an input of any aggregation block includes an output of the input block and an output of each aggregation block before the aggregation block, and an input of the output block includes the output of the input block and an output of the predetermined number of aggregation blocks. Based on a neural network structure search, a feature selection strategy and a feature fusion strategy of the feature fusion framework are determined, where the feature selection strategy includes feature selection operations of the predetermined number of aggregation blocks and the output block, and the feature fusion strategy includes feature fusion operations of the predetermined number of aggregation blocks and the output block. Based on the determined feature selection strategy and the feature fusion strategy, an automatically designed topology is obtained.

Description

Technical Field

[0001] This disclosure generally relates to the field of graph neural network design, and more specifically, to a method and system for designing the topology of graph neural networks. Background Technology

[0002] Graphs are a ubiquitous data model in the real world and scientific fields. Examples include transportation networks, social networks, knowledge graphs, and molecular structures. The research focus is on designing different neural networks based on different graph data to solve different tasks. The predictive power of existing GNN models primarily relies on two modules: aggregation operations and topology. Existing topologies can be divided into two main categories: deep graph neural networks that sequentially stack aggregation operations and shallow graph neural networks that stack multiple aggregation operations in parallel. The former leads to over-smoothing problems, while the latter requires significant computational resources when accessing information about distant neighbors. Summary of the Invention

[0003] This disclosure proposes a method and system for designing the topology of graph neural networks, which can at least solve the above-mentioned problems, or may not solve the above-mentioned problems.

[0004] According to one aspect of this disclosure, a method for designing the topology of a graph neural network is provided. The method includes: acquiring graph data; inputting the graph data into a directed acyclic feature fusion framework, wherein the feature fusion framework includes an input block, a predetermined number of ordered aggregation blocks, and an output block, wherein the input of any aggregation block includes the output of the input block and the outputs of each preceding aggregation block, and the input of the output block includes the output of the input block and the outputs of the predetermined number of aggregation blocks; determining a feature selection strategy and a feature fusion strategy for the feature fusion framework based on neural network structure search, wherein the feature selection strategy includes feature selection operations for the predetermined number of aggregation blocks and the output block, and the feature fusion strategy includes feature fusion operations for the predetermined number of aggregation blocks and the output block; and obtaining the automatically designed topology based on the determined feature selection strategy and the feature fusion strategy.

[0005] Optionally, the graph data may include academic network graph data and / or commodity network graph data.

[0006] Optionally, the node features of the academic network graph data may include bag-of-words feature vectors based on paper keywords, and the edges of the academic network graph data may represent citation relationships between papers. Similarly, the node features of the product network graph data may include bag-of-words feature vectors based on product attributes, and the edges of the product network graph data may represent co-purchase relationships between products.

[0007] Optionally, the operation of the input block may include preprocessing operations to support subsequent operations, the operation of the aggregation block may include a feature fusion operation, an aggregation operation, and at least one feature selection operation, and the operation of the output block may include a predetermined number of features selection operations and a feature fusion operation, wherein the feature selection operation is used to perform feature selection, the feature fusion operation is used to integrate the results of feature selection, and the aggregation operation is used to aggregate information from neighboring nodes.

[0008] Optionally, the step of determining the feature selection strategy and feature fusion strategy of the feature fusion framework based on neural network structure search may include: setting a search space, wherein the search space includes a set of feature selection operations and a set of feature fusion operations; searching for each feature selection operation and feature fusion operation of the predetermined number of aggregation blocks and the output block in the search space based on a differentiable search algorithm; and determining the feature selection strategy and feature fusion strategy of the feature fusion framework based on each searched feature selection operation and feature fusion operation.

[0009] Optionally, the feature selection operation set may include selection operations and non-selection operations, and the feature fusion operation set may include at least one of summation operation, average operation, maximum value operation, minimum value operation, concatenation operation, long short-term memory unit operation and attention mechanism operation.

[0010] Optionally, the differentiable search algorithm may include a relaxation function, which is used to relax the discrete feature selection operation into a continuous one, thereby obtaining the result of the ordered feature selection operation, wherein the relaxation function is expressed as follows:

[0011]

[0012] Where |O| represents the number of operations in the operation set, c k Let α represent the weight of the k-th operation. k Indicates c k The corresponding learnable parameter, λ, represents the temperature parameter used to mitigate the optimization interval.

[0013] Optionally, the topology design method may further include: determining the aggregation operation strategy of the feature fusion framework based on neural network structure search, wherein the aggregation operation strategy includes the aggregation operation of the predetermined number of aggregation blocks; obtaining the automatically designed aggregation operation based on the determined aggregation operation strategy; and obtaining the automatically designed graph neural network based on the automatically designed topology and the aggregation operation.

[0014] Optionally, the search space may further include a set of aggregation operations, and the step of determining the aggregation operation strategy of the feature fusion framework based on the neural network structure search may include: searching for each aggregation operation of the predetermined number of aggregation blocks in the search space based on a differentiable search algorithm; and determining the aggregation operation strategy based on each aggregation operation found.

[0015] Optionally, the set of aggregation operations may include at least one of the following: aggregation operations based on graph convolutional networks, aggregation operations based on graph attention networks, aggregation operations based on graph isomorphic networks, and aggregation operations based on graph samples and aggregation.

[0016] Optionally, the topology design method may further include: acquiring paper domain data based on the academic network graph data and / or product category data based on the product network graph data, based on the automatically designed graph neural network.

[0017] According to another aspect of this disclosure, a topology design system for a graph neural network is provided, wherein the topology design system includes: a data acquisition unit configured to acquire graph data; a unified framework unit configured to input the graph data into a directed acyclic feature fusion framework, wherein the feature fusion framework includes an input block, a predetermined number of ordered aggregation blocks, and an output block, wherein the input of any aggregation block includes the output of the input block and the output of each aggregation block preceding it, and the input of the output block includes the output of the input block and the outputs of the predetermined number of aggregation blocks; a strategy selection unit configured to determine a feature selection strategy and a feature fusion strategy of the feature fusion framework based on a neural network structure search, wherein the feature selection strategy includes feature selection operations of the predetermined number of aggregation blocks and the output block, and the feature fusion strategy includes feature fusion operations of the predetermined number of aggregation blocks and the output block; and a topology design unit configured to acquire the automatically designed topology based on the determined feature selection strategy and the feature fusion strategy.

[0018] Optionally, the graph data may include academic network graph data and / or commodity network graph data.

[0019] Optionally, the node features of the academic network graph data may include bag-of-words feature vectors based on paper keywords, and the edges of the academic network graph data may represent citation relationships between papers. Similarly, the node features of the product network graph data may include bag-of-words feature vectors based on product attributes, and the edges of the product network graph data may represent co-purchase relationships between products.

[0020] Optionally, the operation of the input block may include preprocessing operations to support subsequent operations, the operation of the aggregation block may include a feature fusion operation, an aggregation operation, and at least one feature selection operation, and the operation of the output block may include a predetermined number of features selection operations and a feature fusion operation, wherein the feature selection operation is used to perform feature selection, the feature fusion operation is used to integrate the results of feature selection, and the aggregation operation is used to aggregate information from neighboring nodes.

[0021] Optionally, the strategy selection unit can be configured to: set a search space, wherein the search space includes a set of feature selection operations and a set of feature fusion operations; search for each feature selection operation and feature fusion operation of the predetermined number of aggregation blocks and the output block in the search space based on a differentiable search algorithm; and determine the feature selection strategy and feature fusion strategy of the feature fusion framework based on each feature selection operation and feature fusion operation found.

[0022] Optionally, the feature selection operation set may include selection operations and non-selection operations, and the feature fusion operation set may include at least one of summation operation, average operation, maximum value operation, minimum value operation, concatenation operation, long short-term memory unit operation and attention mechanism operation.

[0023] Optionally, the differentiable search algorithm may include a relaxation function, which is used to relax the discrete feature selection operation into a continuous one, thereby obtaining the result of the ordered feature selection operation, wherein the relaxation function is expressed as follows:

[0024]

[0025] Where |O| represents the number of operations in the operation set, c k Let α represent the weight of the k-th operation. k Indicates c k The corresponding learnable parameter, λ, represents the temperature parameter used to mitigate the optimization interval.

[0026] Optionally, the topology design system may further include: an aggregation design unit configured to: determine the aggregation operation strategy of the feature fusion framework based on neural network structure search, wherein the aggregation operation strategy includes the aggregation operation of the predetermined number of aggregation blocks; obtain the automatically designed aggregation operation based on the determined aggregation operation strategy; and obtain the automatically designed graph neural network based on the automatically designed topology and the aggregation operation.

[0027] Optionally, the search space may further include a set of aggregation operations, and the aggregation design unit may be configured to: search for each aggregation operation of the predetermined number of aggregation blocks in the search space based on a differentiable search algorithm; and determine the aggregation operation strategy based on each aggregation operation found.

[0028] Optionally, the set of aggregation operations may include at least one of the following: aggregation operations based on graph convolutional networks, aggregation operations based on graph attention networks, aggregation operations based on graph isomorphic networks, and aggregation operations based on graph samples and aggregation.

[0029] Optionally, the topology design system may further include: a data processing unit configured to: acquire paper domain data based on the academic network graph data and / or product category data based on the product network graph data, based on the automatically designed graph neural network.

[0030] According to another aspect of this disclosure, a system is provided that includes at least one computing device and at least one storage device for storing instructions, wherein the instructions, when executed by the at least one computing device, cause the at least one computing device to perform a graph neural network topology design method according to this disclosure.

[0031] According to another aspect of this disclosure, a computer-readable storage medium for storing instructions is provided, wherein when the instructions are executed by at least one computing device, the at least one computing device causes the at least one computing device to perform a graph neural network topology design method according to this disclosure.

[0032] The graph neural network topology design method and system according to exemplary embodiments of the present disclosure can adaptively utilize the features of different layers. It can not only extract local neighbor information while acquiring higher-level features, but also improve the performance of graph neural networks while saving computing resources, especially greatly alleviating the oversmoothing problem. Attached Figure Description

[0033] These and / or other aspects and advantages of the invention will become clear and more readily understood from the following description of the embodiments, taken in conjunction with the accompanying drawings, wherein:

[0034] Figure 1This is a flowchart illustrating a method for designing the topology of a graph neural network according to exemplary embodiments of the present disclosure;

[0035] Figure 2 This is a schematic diagram illustrating a feature fusion framework according to an exemplary embodiment of the present disclosure;

[0036] Figure 3 This illustrates exemplary embodiments according to this disclosure. Figure 1 Flowchart of step S102;

[0037] Figure 4 This is a block diagram illustrating a graph neural network topology design system according to exemplary embodiments of the present disclosure;

[0038] Figure 5 This is a schematic diagram illustrating the performance improvement effect of a model according to an exemplary embodiment of the present disclosure. Detailed Implementation

[0039] The following description, provided with reference to the accompanying drawings, is intended to aid in a full understanding of embodiments of the invention as defined by the claims and their equivalents. Various specific details are included to aid understanding, but these details are to be considered exemplary only. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Furthermore, for clarity and brevity, descriptions of well-known functions and structures are omitted.

[0040] It should be noted that the phrase "at least one of several items" in this disclosure refers to three parallel cases: "any one of the several items", "a combination of any number of the several items", and "all of the several items". For example, "including at least one of A and B" includes the following three parallel cases: (1) including A; (2) including B; (3) including A and B. As another example, "performing at least one of step one and step two" indicates the following three parallel cases: (1) performing step one; (2) performing step two; (3) performing both step one and step two.

[0041] In recent years, Generative Neural Networks (GNNs) have demonstrated superior performance in various practical applications. For further improvements to GNN performance, besides designing aggregation operations to better aggregate neighbor information, topology design is also crucial. Generally, existing GNN topology designs focus on two aspects: First, stacking aggregation operations to obtain higher-level features; however, this approach easily leads to oversmoothing, causing performance degradation. Second, using multiple aggregation operations in each layer to independently extract and fuse features to enhance the extraction of local neighbor information; while this approach alleviates oversmoothing and provides rich local features, it requires significant computational resources. Therefore, exploring how to leverage strengths and mitigate weaknesses to design the optimal GNN topology is a work of practical significance.

[0042] Existing GNN models are designed based on expert experience, but different tasks require different GNN design preferences. For example, current automated machine learning methods, such as those proposed in the papers "Search to aggregate neighborhood for graph neural network" and "GraphNAS: Graph Neural Architecture Search with Reinforcement," mostly design GNNs based on stacked aggregation operations, which cannot fully cover existing GNN solutions. Furthermore, the method proposed in the paper "DON'T STACK LAYERS IN GRAPH NEURAL NETWORKS, WIRE THEMRANDOMLY" only considers the connections between aggregation operations and cannot fully describe the topological structure of GNNs. Moreover, by comparing existing GNN models, it can be observed that, in addition to differences in aggregation function design, feature selection and fusion strategies lead to the main differences between GNNs.

[0043] Therefore, this disclosure uses a unified feature fusion framework to encompass existing topology design methods, unifying existing topology design with feature selection and fusion strategies, thereby transforming GNN topology design into feature selection and fusion strategy selection. Furthermore, based on the aforementioned feature fusion framework, this disclosure proposes a GNN topology design scheme based on Neural Architecture Search (NAS). This scheme includes an effective search space and a gradient descent-based differentiable search algorithm, which can quickly find a better-performing GNN topology, thus automating the design of GNN topologies. This approach leverages strengths and avoids weaknesses to produce a higher-performing GNN, surpassing the performance of the best GNNs designed by human experts on different datasets.

[0044] Meanwhile, the graph neural network topology design method and system according to the exemplary embodiments of this disclosure can be applied to various GNN-based tasks, such as node classification, graph classification, link prediction and knowledge graph tasks.

[0045] The following reference Figures 1 to 5 A method and system for designing the topology of a graph neural network according to exemplary embodiments of the present disclosure are described in detail.

[0046] Figure 1 This is a flowchart illustrating a method for designing the topology of a graph neural network according to an exemplary embodiment of the present disclosure.

[0047] Reference Figure 1 In step S101, graph data can be acquired. Here, graph data may include academic network graph data and / or product network graph data. Specifically, graph data can be represented as G = (A, H), where... It can represent an adjacency matrix, where N can represent the number of nodes. The feature matrix can represent the node, and d can represent the feature dimension. More specifically, the node features of academic network graph data can include bag-of-words feature vectors based on paper keywords, and the edges of academic network graph data can represent the citation relationships between papers. Similarly, the node features of product network graph data can include bag-of-words feature vectors based on product attributes, and the edges of product network graph data can represent the co-purchase relationships between products. As an example, in the academic network Cora, each paper is abstracted as a node, and the node's features are bag-of-words feature vectors generated based on the keywords in the paper. The edges between nodes represent the citation relationships between papers, and the field of the paper can be determined based on the node's features and citation relationships. In the product network Computer, each node represents a product, and the node's features are bag-of-words feature vectors generated based on product attributes. If there is an edge between two nodes, it means that the two products have been co-purchased, and the product type can be determined based on product attributes and the purchasing characteristics between products.

[0048] Next, in step S102, the graph data can be input into a directed acyclic feature fusion framework. Here, existing topologies include stacked aggregation operation structures and multiple aggregation operation structures. These existing topologies can be unified into a single feature fusion framework, encompassing both stacked and multiple aggregation operation structures. This unifies topology designs based on different feature selection and fusion strategies. Then, graph data from practical applications such as academic or commercial networks is input into the feature fusion framework, and topology design is performed based on this framework. Without loss of generality, the feature fusion framework can be represented as a directed acyclic graph (DAG), composed of ordered sequences of blocks.

[0049] According to an exemplary embodiment of this disclosure, the feature fusion framework may include an input block, a predetermined number N ordered aggregation blocks, and an output block. Here, the input to any aggregation block may include the output of the input block and the outputs of each aggregation block preceding it, and the input to the output block may include the output of the input block and the outputs of the predetermined number N aggregation blocks. Furthermore, the predetermined number N can be set by those skilled in the art according to actual circumstances.

[0050] According to exemplary embodiments of this disclosure, the operations on the input block may include preprocessing operations to support subsequent operations; the operations on the aggregation block may include a feature fusion operation, an aggregation operation, and at least one feature selection operation; and the operations on the output block may include a predetermined number plus one (i.e., N+1) feature selection operations and a feature fusion operation. Here, the feature selection operation can be used to perform feature selection, the feature fusion operation can be used to integrate the results of feature selection, the aggregation operation can be used to aggregate information from neighboring nodes, and the number of feature selection operations for any aggregation block depends on the number of aggregation blocks preceding that aggregation block. Specifically, in any aggregation block, for each feature input to that aggregation block, a feature selection operation is performed, thereby processing each feature. Further, the preprocessing operations of the input block can be implemented using a Multilayer Perceptron (MLP); the output block can also be further post-processed using an MLP on the information after the N+1 feature selection operations and the feature fusion operation, thereby improving the model's capabilities. See below for reference. Figure 2 A feature fusion framework according to exemplary embodiments of the present disclosure is described in detail.

[0051] Figure 2 This is a schematic diagram illustrating a feature fusion framework according to an exemplary embodiment of the present disclosure.

[0052] Reference Figure 2 The feature fusion framework can be represented as a DAG, consisting of an input block and N aggregation blocks (as an example). Figure 2 (a) In the unified framework, N is 4) and there is one output block. Input block 0 may include an MLP for performing simple preprocessing operations to support subsequent blocks. The aggregation block includes feature selection, feature fusion, and aggregation operations. For Figure 2 (b) For the aggregate block i shown in the aggregate block diagram, based on the output results of the previous i blocks, i feature selection operations f can be used. s This is how it's processed. Specifically, for aggregate block 4, four feature selection operations f can be used. s The outputs of input block 0, aggregation block 1, aggregation block 2, and aggregation block 3 are processed. Then, a feature fusion operation f can be performed. f The results of feature selection are then integrated, and subsequently, an aggregation operation f can be performed. a Information is aggregated from neighbors. Additionally, in output block 5, five feature selection operations are performed on the outputs of input block 0, aggregation block 1, aggregation block 2, aggregation block 3, and aggregation block 4, followed by a feature fusion operation. Then, an MLP is used to further post-process the information after the five feature selection operations and the feature fusion operation. Based on this, the output feature H of aggregation block i can be represented by equation (1).i :

[0053]

[0054] Return to reference Figure 1 In step S103, based on the neural network architecture search, the feature selection strategy and feature fusion strategy of the feature fusion framework can be determined. Here, the feature selection strategy may include feature selection operations for the output block and a predetermined number of aggregate blocks, and the feature fusion strategy may include feature fusion operations for the output block and a predetermined number of aggregate blocks. Generally, the neural network architecture search method uses a search algorithm to automatically find a state-of-the-art (SOTA) neural network in a predefined search space. (See below for reference.) Figure 3 Description of exemplary embodiments according to this disclosure Figure 1 Step S102.

[0055] Figure 3 This illustrates exemplary embodiments according to this disclosure. Figure 1 The flowchart for step S102.

[0056] Reference Figure 3 In step S301, a search space can be set. Here, the search space may include the feature selection operation set O. s and feature fusion operation set O f .

[0057] According to an exemplary embodiment of this disclosure, the feature selection operation set may include selection operations and non-selection operations. Specifically, the selection operation may be represented as IDENTITY, the non-selection operation may be represented as ZERO, and the selection operation may be represented by equation (2), and the non-selection operation may be represented by equation (3):

[0058] f(h)=h (2)

[0059] f(h) = 0·h (3)

[0060] Here, h can represent the eigenvector.

[0061] According to exemplary embodiments of this disclosure, the feature fusion operation set may include, but is not limited to, at least one of the following: summation operation, averaging operation, maximizing operation, minimizing operation, concatenation operation, long short-term memory unit operation, and attention mechanism operation. Furthermore, the operations in the feature fusion operation set can be configured by those skilled in the art according to actual conditions. Specifically, the summation operation may be represented as SUM, the averaging operation as MEAN, the maximizing operation as MAX, the minimizing operation as MIN, the concatenation operation as CONCAT, the long short-term memory unit operation as LSTM, and the attention mechanism operation as ATT.

[0062] Next, in step S302, a differentiable search algorithm can be used to search the search space for each feature selection and feature fusion operation of a predetermined number of aggregation blocks and output blocks. By using a differentiable search algorithm, the efficiency and effectiveness of the search can be guaranteed.

[0063] According to an exemplary embodiment of this disclosure, in a differentiable search algorithm, a supernet can be defined to encompass all models based on the aforementioned feature fusion framework and search space, and this is achieved through mixing the candidate operations. The result of the mixing candidate operations can be calculated by weighted summation of all candidate operations, and can be expressed by equation (4):

[0064]

[0065] Here, x can represent the feature for the mixed candidate operation, and c k ∈(0,1) can represent the k-th candidate operation o k The weights of (·) are given by |O|, which can represent the number of operations in the operation set. Generally, the operation weights c k It can be represented by a relaxation function, which can be expressed by equation (5):

[0066]

[0067] Here, α k Can be represented as C k The corresponding learnable supernet parameter.

[0068] Based on relaxation functions, discrete feature selection operations can be relaxed into continuous operations, thereby gradually generating the final result in the output block. Furthermore, gradient descent can be used to optimize the supernet, significantly accelerating the search process. After completing the search, the operation with the largest weight in each mixed candidate operation can be retained, ultimately yielding the searched GNN.

[0069] Although the supernet can be optimized during the search process and then the GNN can be derived after the search is complete, it is difficult to generate the best subnet from the best supernet because only the childnet is selected in the end. This is called the optimization gap. Specifically, since the feature selection operation set includes two opposite operations, namely ZERO and IDENTITY, the performance degradation caused by the optimization gap based on the feature fusion framework is very significant.

[0070] Generally, for the features generated by block i, the result of the hybrid selection operation in block j (j>i) can be expressed as a weighted sum of the ZERO and IDENTITY operations. Here, the hybrid selection operation can be represented by equation (6):

[0071]

[0072] here, x can represent the mixed selection operation of features generated from block i in block j. i It can represent the feature generated by block i and the mixed selection operation performed in block j.

[0073] Next, the result of the mixing and fusion operation in block j can be expressed by equation (7):

[0074]

[0075] here, This can represent the k-th feature fusion operation in block j.

[0076] Generally, when the weight c1 of the ZERO operation is greater than the weight c2 of the IDENTITY operation, in the subnet, for equation (6), a ZERO tensor result is expected through the ZERO operation; simultaneously, in equation (7), if no features are selected in the block, a ZERO tensor result is also expected. However, in the supernet, for equation (6), the result of the mixed operation of the ZERO operation is... In other words, in this case, for each hybrid selection operation, when expecting to obtain a ZERO tensor result, what is actually obtained is... It is evident that in equation (6), the IDENTITY operation has a significant impact on the ZERO operation, and this impact accumulates as the feature selection operation in the feature fusion framework proceeds. Since the feature selection operation set includes two opposite operations, ZERO and IDENTITY, the difference between the supernet and subnet results in the feature fusion framework is very significant. Furthermore, this difference prevents the optimal subnet from being derived from the optimal supernet, resulting in a significant performance degradation due to the optimization interval based on the feature fusion framework.

[0077] Therefore, considering the impact of the IDENTITY operation, a temperature parameter can be added to the relaxation function. Specifically, differentiable search algorithms may include a relaxation function, which is used to relax discrete feature selection operations into continuous ones, thereby obtaining the results of ordered feature selection operations. Here, the relaxation function can be represented by equation (8):

[0078]

[0079] Here, |O| can represent the number of operations in the operation set, c k Let α represent the weight of the k-th operation. k Can be represented as c k The corresponding learnable parameter, λ, can represent the temperature parameter used to mitigate the optimization interval.

[0080] By adding a small λ to the relaxation function, the operation weight vector c is closer to the one-hot vector, and for equation (6), the result of the mixed selection operation is closer to the ZERO tensor when performing the ZERO operation. Here, the value of λ can be set to 0.001, or it can be set by those skilled in the art according to the actual situation.

[0081] Next, in step S303, the feature selection strategy and feature fusion strategy of the feature fusion framework can be determined based on each feature selection operation and feature fusion operation found.

[0082] Return to reference Figure 1 In step S104, an automatically designed topology can be obtained based on the determined feature selection and feature fusion strategies. Further, aggregation operations can be selected, and then the selected aggregation operations can be used in each aggregation block. For example, an aggregation operation based on Graph Sample and Aggregate (GraphSAGE) can be used in each aggregation block, thereby obtaining a newly designed graph neural network based on the automatically designed topology and the selected aggregation operations.

[0083] According to exemplary embodiments of this disclosure, the search space described above may further include an aggregation operation set O. a Specifically, the aggregation operation set may include, but is not limited to, at least one of the following: aggregation operations based on Graph Convolution Network (GCN), aggregation operations based on Graph Attention Network (GAT), aggregation operations based on Graph Isomorphism Network (GIN), and aggregation operations based on GraphSAGE. Furthermore, the operations in the aggregation operation set can be configured by those skilled in the art according to the actual situation. Specifically, aggregation operations based on GCN can be represented as GCN, aggregation operations based on GAT can be represented as GAT, aggregation operations based on GIN can be represented as GIN, and aggregation operations based on GraphSAGE can be represented as SAGE.

[0084] Building upon this, as an example, the aggregation operation strategy of the feature fusion framework can also be determined based on neural network structure search. Here, the aggregation operation strategy includes aggregation operations for a predetermined number of aggregation blocks. Specifically, a differentiable search algorithm can be used to search for each aggregation operation of the predetermined number of aggregation blocks in the search space; then, an aggregation operation strategy can be determined based on each found aggregation operation. Next, based on the determined aggregation operation strategy, automatically designed aggregation operations can be obtained. Following that, based on the automatically designed topology and aggregation operations, an automatically designed graph neural network can be obtained.

[0085] According to exemplary embodiments of this disclosure, paper domain data based on academic network graph data and / or product category data based on product network graph data can be obtained based on automatically designed graph neural networks.

[0086] The graph neural network topology design method according to exemplary embodiments of the present disclosure can adaptively utilize the features of different layers, and can make up for the shortcomings of existing topology design methods while enjoying the benefits of existing topology design methods, especially greatly alleviating the oversmoothing problem.

[0087] Figure 4 This is a block diagram illustrating a graph neural network topology design system according to exemplary embodiments of the present disclosure.

[0088] Reference Figure 4 The graph neural network topology design system 400 according to an exemplary embodiment of the present disclosure may include a data acquisition unit 401, a unified framework unit 402, a strategy selection unit 403, and a topology design unit 404.

[0089] The data acquisition unit 401 can acquire graph data. Here, the graph data may include academic network graph data and / or product network graph data. Specifically, the node features of academic network graph data may include bag-of-words feature vectors based on paper keywords, and the edges of academic network graph data may represent citation relationships between papers; and the node features of product network graph data may include bag-of-words feature vectors based on product attributes, and the edges of product network graph data may represent co-purchase relationships between products.

[0090] The unified framework unit 402 can input graph data into a directed acyclic feature fusion framework. Here, the feature fusion framework can consist of an ordered sequence of blocks, including an input block, a predetermined number N ordered aggregation blocks, and an output block. Furthermore, the input to any aggregation block can include the output of the input block and the outputs of each preceding aggregation block, and the input to the output block can include the output of the input block and the outputs of the predetermined number N aggregation blocks. Even further, the predetermined number N can be set by those skilled in the art according to actual conditions.

[0091] According to exemplary embodiments of this disclosure, the operations of the input block may include preprocessing operations to support subsequent operations, the operations of the aggregation block may include a feature fusion operation, an aggregation operation, and at least one feature selection operation, and the operations of the output block may include a predetermined number plus one (i.e., N+1) of feature selection operations and a feature fusion operation. Here, the feature selection operation can be used to perform feature selection, the feature fusion operation can be used to integrate the results of feature selection, and the aggregation operation can be used to aggregate information from neighboring nodes.

[0092] The strategy selection unit 403 can determine the feature selection strategy and feature fusion strategy of the feature fusion framework based on neural network structure search. Here, the feature selection strategy may include feature selection operations of the output block and a predetermined number of aggregate blocks, and the feature fusion strategy may include feature fusion operations of the output block and a predetermined number of aggregate blocks.

[0093] According to an exemplary embodiment of this disclosure, the strategy selection unit 402 may set a search space. Here, the search space may include a feature selection operation set O. s and feature fusion operation set O f Furthermore, the feature selection operation set may include selection operations and non-selection operations. Meanwhile, the feature fusion operation set may include, but is not limited to, at least one of the following: summation operation, averaging operation, maximum value operation, minimum value operation, concatenation operation, long short-term memory unit operation, and attention mechanism operation. Moreover, the operations in the feature fusion operation set can be configured by those skilled in the art according to the actual situation.

[0094] According to an exemplary embodiment of this disclosure, the strategy selection unit 402 may search in the search space for each feature selection operation and feature fusion operation of a predetermined number of aggregation blocks and output blocks based on a differentiable search algorithm. Here, the differentiable search algorithm may include a relaxation function, which is used to relax discrete feature selection operations into continuous ones, thereby obtaining the result of ordered feature selection operations. Further, the relaxation function can be represented by equation (8) as described above.

[0095] According to an exemplary embodiment of the present disclosure, the strategy selection unit 402 can determine the feature selection strategy and feature fusion strategy of the feature fusion framework based on each feature selection operation and feature fusion operation found.

[0096] The topology design unit 404 can obtain the topology structure for automated design based on a defined feature selection strategy and feature fusion strategy.

[0097] According to exemplary embodiments of this disclosure, the search space described above may further include an aggregation operation set O. aHere, the set of aggregation operations may include, but is not limited to, at least one of GCN-based aggregation operations, GAT-based aggregation operations, GIN-based aggregation operations, and GraphSAGE-based aggregation operations. Furthermore, the operations in the set of aggregation operations can be configured by those skilled in the art according to the actual situation.

[0098] Based on this, the graph neural network topology design system 400 according to the exemplary embodiments of this disclosure may further include an aggregation design unit and a data processing unit.

[0099] According to an exemplary embodiment of this disclosure, the aggregation design unit can determine the aggregation operation strategy of the feature fusion framework based on neural network structure search. Here, the aggregation operation strategy includes aggregation operations of a predetermined number of aggregation blocks. Specifically, the aggregation design unit can search for each aggregation operation of the predetermined number of aggregation blocks in the search space based on a differentiable search algorithm; then, the aggregation design unit can determine the aggregation operation strategy based on each searched aggregation operation. Next, the aggregation design unit can obtain automatically designed aggregation operations based on the determined aggregation operation strategy. Following that, the aggregation design unit can obtain an automatically designed graph neural network based on the automatically designed topology and aggregation operations.

[0100] According to an exemplary embodiment of this disclosure, the data processing unit can acquire paper domain data based on the academic network graph data and / or product category data based on the product network graph data, based on the automatically designed graph neural network.

[0101] The graph neural network topology design method and system according to exemplary embodiments of this disclosure can adaptively utilize features from different layers. This not only allows for the extraction of local neighbor information while acquiring higher-level features, but also improves the performance of the graph neural network while conserving computational resources, especially significantly mitigating the oversmoothing problem. See below for reference. Figure 5 The performance improvement of the model according to the exemplary embodiments of this disclosure is described in detail.

[0102] Figure 5 This is a schematic diagram illustrating the performance improvement effect of a model according to an exemplary embodiment of the present disclosure.

[0103] Reference Figure 5 As an example, the aggregation operation strategy for the model in the "SAGE" section uses GraphSAGE-based aggregation operations; the aggregation operation strategy for the model in the "GAT" section uses GAT-based aggregation operations; and the aggregation operation strategy for the model in the "Learnable" section is automatically designed according to exemplary embodiments of this disclosure. Furthermore, for model F... 2 SAGE, F 2GAT and F 2 The GNN's topology is automatically designed according to exemplary embodiments of this disclosure, while the topologies of other models are existing designs. As shown in the figure, based on the automatically designed topologies according to exemplary embodiments of this disclosure, for academic networks Cora, PubMed, DBLP, and Physics, and the commodity network Computer, model F... 2 SAGE, F 2 GAT and F 2 GNN achieved optimal results, ranking first in performance comparisons across its respective components. Compared to existing GNNs, it offers at least a 2% improvement in accuracy, and eliminates the need for expert-designed GNN structures for feature datasets, saving time and effort. Furthermore, the GNN with automated topology design based on exemplary embodiments of this disclosure is universally applicable and can be widely used in various graph data. Moreover, the automated design of GNN topologies based on exemplary embodiments of this disclosure provides excellent scalability, allowing for easy application to other domains where GNNs can be used, and achieving significant performance improvements on other types of graph data.

[0104] The above has been referred to Figures 1 to 5 A method and system for designing the topology of a graph neural network according to exemplary embodiments of the present disclosure are described.

[0105] Figure 4 The units in the illustrated graph neural network topology design system can be configured as software, hardware, firmware, or any combination thereof to perform specific functions. For example, each unit may correspond to a dedicated integrated circuit, pure software code, or a module combining software and hardware. Furthermore, one or more functions implemented by each unit can also be uniformly executed by components in a physical entity device (e.g., a processor, client, or server).

[0106] In addition, refer to Figure 1 The described method for designing the topology of a graph neural network can be implemented by a program (or instructions) recorded on a computer-readable storage medium. For example, according to an exemplary embodiment of this disclosure, a computer-readable storage medium may be provided to store instructions, wherein when the instructions are executed by at least one computing device, the at least one computing device causes the at least one computing device to perform a method for assisting human text annotation according to this disclosure.

[0107] The computer program in the aforementioned computer-readable storage medium can run in an environment deployed in computer devices such as clients, hosts, agent devices, and servers. It should be noted that the computer program can also be used to perform additional steps beyond those described above, or to perform more specific processing while performing the above steps. The details of these additional steps and further processing are already described in the reference... Figure 1 The relevant methods were mentioned in the description of the process, so they will not be repeated here to avoid repetition.

[0108] It should be noted that each unit in the graph neural network topology design system according to the exemplary embodiments of this disclosure can rely entirely on the operation of a computer program to realize its corresponding function. That is, each unit corresponds to each step in the functional architecture of the computer program, so that the entire system is called through a special software package (e.g., a lib library) to realize its corresponding function.

[0109] on the other hand, Figure 4 The units shown can also be implemented using hardware, software, firmware, middleware, microcode, or any combination thereof. When implemented in software, firmware, middleware, or microcode, the program code or code segment used to perform the corresponding operation can be stored in a computer-readable medium such as a storage medium, so that the processor can perform the corresponding operation by reading and running the corresponding program code or code segment.

[0110] For example, exemplary embodiments of the present disclosure can also be implemented as a computing device including a storage component and a processor, wherein the storage component stores a set of computer-executable instructions, and when the set of computer-executable instructions is executed by the processor, a graph neural network topology design method according to exemplary embodiments of the present disclosure is executed.

[0111] Specifically, the computing device can be deployed on a server or client, or on node devices in a distributed network environment. Furthermore, the computing device can be a PC, tablet, personal digital assistant, smartphone, web application, or other device capable of executing the aforementioned set of instructions.

[0112] Here, the computing device is not necessarily a single computing device, but can be any collection of devices or circuits capable of executing the aforementioned instructions (or instruction sets) individually or in combination. The computing device can also be part of an integrated control system or system manager, or can be configured to interconnect with a portable electronic device locally or remotely (e.g., via wireless transmission) through an interface.

[0113] In a computing device, a processor may include a central processing unit (CPU), a graphics processing unit (GPU), a programmable logic device, a dedicated processor system, a microcontroller, or a microprocessor. By way of example and not limitation, a processor may also include an analog processor, a digital processor, a microprocessor, a multi-core processor, a processor array, a network processor, etc.

[0114] Some operations described in the graph neural network topology design method according to exemplary embodiments of the present disclosure can be implemented in software, some operations can be implemented in hardware, and some operations can be implemented in a combination of software and hardware.

[0115] The processor can execute instructions or code stored in one of the storage components, which can also store data. Instructions and data can also be sent and received over a network via a network interface device, which can employ any known transport protocol.

[0116] Storage components can be integrated with the processor, for example, by placing RAM or flash memory within an integrated circuit microprocessor. Alternatively, storage components can include separate devices, such as external disk drives, storage arrays, or other storage devices that can be used by any database system. Storage components and the processor can be operatively coupled, or can communicate with each other, for example, via I / O ports, network connections, etc., enabling the processor to read files stored in the storage component.

[0117] In addition, the computing device may include a video display (such as a liquid crystal display) and a user interaction interface (such as a keyboard, mouse, touch input device, etc.). All components of the computing device may be interconnected via a bus and / or network.

[0118] The topology design method for graph neural networks according to exemplary embodiments of this disclosure can be described as various interconnected or coupled functional blocks or functional diagrams. However, these functional blocks or functional diagrams can be equally integrated into a single logic device or operate according to non-precise boundaries.

[0119] Therefore, refer to Figure 1 The described graph neural network topology design method can be implemented through a system including at least one computing device and at least one storage device for storing instructions.

[0120] According to an exemplary embodiment of the present disclosure, at least one computing device is a computing device for executing a topology design method for a graph neural network according to an exemplary embodiment of the present disclosure. A storage device stores a set of computer-executable instructions. When the set of computer-executable instructions is executed by the at least one computing device, a reference is executed. Figure 1 The described method for designing the topology of graph neural networks.

[0121] The foregoing has described various exemplary embodiments of this disclosure. It should be understood that the foregoing description is exemplary only and not exhaustive, and this disclosure is not limited to the disclosed exemplary embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A classification method based on graph neural networks, wherein, The classification method includes: Acquire graph data, wherein the graph data includes academic network graph data and / or product network graph data, wherein the node features of the academic network graph data include bag-of-words feature vectors based on paper keywords, and the edges of the academic network graph data represent citation relationships between papers; and the node features of the product network graph data include bag-of-words feature vectors based on product attributes, and the edges of the product network graph data represent co-purchase relationships between products. The graph data is input into a directed acyclic feature fusion framework, wherein the feature fusion framework includes an input block, a predetermined number of ordered aggregation blocks, and an output block. The input of any aggregation block includes the output of the input block and the outputs of each aggregation block preceding it. The input of the output block includes the output of the input block and the outputs of the predetermined number of aggregation blocks. Based on neural network structure search, the feature selection strategy and feature fusion strategy of the feature fusion framework are determined, wherein the feature selection strategy includes feature selection operations of the predetermined number of aggregation blocks and the output block, and the feature fusion strategy includes feature fusion operations of the predetermined number of aggregation blocks and the output block; Based on the determined feature selection strategy and feature fusion strategy, the topology structure of the automated design is obtained; Based on neural network structure search, the aggregation operation strategy of the feature fusion framework is determined, wherein the aggregation operation strategy includes the aggregation operation of the predetermined number of aggregation blocks; based on the determined aggregation operation strategy, the automatically designed aggregation operation is obtained; based on the automatically designed topology and the aggregation operation, the automatically designed graph neural network is obtained. Based on the automatically designed graph neural network, paper domain data based on the academic network graph data and / or product category data based on the product network graph data are obtained.

2. The classification method based on graph neural networks as described in claim 1, wherein, The operations on the input block include preprocessing operations to support subsequent operations; the operations on the aggregation block include a feature fusion operation, an aggregation operation, and at least one feature selection operation; and the operations on the output block include a predetermined number of features plus one feature selection operation and one feature fusion operation. The feature selection operation is used to select features, the feature fusion operation is used to integrate the results of feature selection, and the aggregation operation is used to aggregate information from neighboring nodes.

3. The classification method based on graph neural networks as described in claim 2, wherein, The steps for determining the feature selection strategy and feature fusion strategy of the feature fusion framework based on neural network structure search include: A search space is set, wherein the search space includes a feature selection operation set and a feature fusion operation set; Based on a differentiable search algorithm, a search is performed in the search space for each feature selection operation and feature fusion operation of the predetermined number of aggregation blocks and the output block; Based on each feature selection operation and feature fusion operation found, the feature selection strategy and feature fusion strategy of the feature fusion framework are determined.

4. The classification method based on graph neural networks as described in claim 3, wherein, The feature selection operation set includes selection operations and non-selection operations, and the feature fusion operation set includes at least one of the following: summation operation, average operation, maximum value operation, minimum value operation, concatenation operation, long short-term memory unit operation, and attention mechanism operation.

5. The classification method based on graph neural networks as described in claim 4, wherein, The differentiable search algorithm includes a relaxation function, which is used to relax the discrete feature selection operation into a continuous one, thereby obtaining the result of the ordered feature selection operation. The relaxation function is expressed as follows: in, This indicates the number of operations in the operation set. Indicates the first The weight of each operation, Indicates and The corresponding learnable parameters, This indicates the temperature parameter used to mitigate the optimization interval.

6. The classification method based on graph neural networks as described in claim 5, wherein, The search space also includes a set of aggregation operations, wherein the step of determining the aggregation operation strategy of the feature fusion framework based on the neural network structure search includes: Based on a differentiable search algorithm, a search is performed on each aggregation operation of the predetermined number of aggregation blocks in the search space; Based on each aggregation operation found, the aggregation operation strategy is determined.

7. The classification method based on graph neural networks as described in claim 6, wherein, The set of aggregation operations includes at least one of the following: aggregation operations based on graph convolutional networks, aggregation operations based on graph attention networks, aggregation operations based on graph isomorphic networks, and aggregation operations based on graph samples and aggregation.

8. A classification system based on graph neural networks, wherein, The classification system includes: The data acquisition unit is configured to acquire graph data, wherein the graph data includes academic network graph data and / or product network graph data, the node features of the academic network graph data include bag-of-words feature vectors based on paper keywords, the edges of the academic network graph data represent citation relationships between papers, and the node features of the product network graph data include bag-of-words feature vectors based on product attributes, the edges of the product network graph data represent co-purchase relationships between products; A unified framework unit is configured to input the graph data into a directed acyclic feature fusion framework, wherein the feature fusion framework includes an input block, a predetermined number of ordered aggregation blocks, and an output block, wherein the input of any aggregation block includes the output of the input block and the output of each aggregation block preceding it, and the input of the output block includes the output of the input block and the output of the predetermined number of aggregation blocks; The strategy selection unit is configured to determine the feature selection strategy and feature fusion strategy of the feature fusion framework based on neural network structure search, wherein the feature selection strategy includes feature selection operations of the predetermined number of aggregation blocks and the output block, and the feature fusion strategy includes feature fusion operations of the predetermined number of aggregation blocks and the output block; The topology design unit is configured to obtain an automatically designed topology based on the determined feature selection strategy and the feature fusion strategy; An aggregation design unit is configured to determine the aggregation operation strategy of the feature fusion framework based on neural network structure search, wherein the aggregation operation strategy includes aggregation operations of the predetermined number of aggregation blocks; based on the determined aggregation operation strategy, obtain the automatically designed aggregation operations; and based on the automatically designed topology and the aggregation operations, obtain the automatically designed graph neural network. The data processing unit is configured to acquire paper domain data based on the academic network graph data and / or product category data based on the product network graph data, based on the automatically designed graph neural network.

9. The classification system based on graph neural networks as described in claim 8, wherein, The operations on the input block include preprocessing operations to support subsequent operations; the operations on the aggregation block include a feature fusion operation, an aggregation operation, and at least one feature selection operation; and the operations on the output block include a predetermined number of features plus one feature selection operation and one feature fusion operation. The feature selection operation is used to select features, the feature fusion operation is used to integrate the results of feature selection, and the aggregation operation is used to aggregate information from neighboring nodes.

10. The classification system based on graph neural networks as described in claim 9, wherein, The strategy selection unit is configured as follows: A search space is set, wherein the search space includes a feature selection operation set and a feature fusion operation set; Based on a differentiable search algorithm, a search is performed in the search space for each feature selection operation and feature fusion operation of the predetermined number of aggregation blocks and the output block; Based on each feature selection operation and feature fusion operation found, the feature selection strategy and feature fusion strategy of the feature fusion framework are determined.

11. The classification system based on graph neural networks as described in claim 10, wherein, The feature selection operation set includes selection operations and non-selection operations, and the feature fusion operation set includes at least one of the following: summation operation, average operation, maximum value operation, minimum value operation, concatenation operation, long short-term memory unit operation, and attention mechanism operation.

12. The classification system based on graph neural networks as described in claim 11, wherein, The differentiable search algorithm includes a relaxation function, which is used to relax the discrete feature selection operation into a continuous one, thereby obtaining the result of the ordered feature selection operation. The relaxation function is expressed as follows: in, This indicates the number of operations in the operation set. Indicates the first The weight of each operation, Indicates and The corresponding learnable parameters, This indicates the temperature parameter used to mitigate the optimization interval.

13. The classification system based on graph neural networks as described in claim 10, wherein, The search space also includes a set of aggregation operations, wherein the aggregation design unit is configured as follows: Based on a differentiable search algorithm, a search is performed on each aggregation operation of the predetermined number of aggregation blocks in the search space; Based on each aggregation operation found, the aggregation operation strategy is determined.

14. The classification system based on graph neural networks as described in claim 13, wherein, The set of aggregation operations includes at least one of the following: aggregation operations based on graph convolutional networks, aggregation operations based on graph attention networks, aggregation operations based on graph isomorphic networks, and aggregation operations based on graph samples and aggregation.

15. A system comprising at least one computing device and at least one storage device for storing instructions, wherein, When the instruction is executed by the at least one computing device, it causes the at least one computing device to perform the graph neural network-based classification method as described in any one of claims 1 to 7.

16. A computer-readable storage medium for storing instructions, wherein, When the instruction is executed by at least one computing device, it causes the at least one computing device to perform the graph neural network-based classification method as described in any one of claims 1 to 7.

Images