A few-shot graph classification method and system based on graphical decoupled meta-learning

Abstract

A few-shot graph classification method and system based on graphical decoupling meta-learning. The method comprises: explicitly modeling the many-to-many relationship between molecules and attributes by constructing a multi-relation molecular graph (MMG), and separating task-specific and task-invariant features by using a decoupling graph learning strategy, and combining a meta-learning framework to realize molecular attribute prediction in a few-shot scenario. The scheme improves the sample prediction accuracy and enhances the interpretability of the features.

Description

Technical Field

[0001] This invention belongs to the field of molecular attribute prediction, and specifically relates to a few-shot graph classification method and system based on graph decoupling meta-learning. Background Technology

[0002] Molecular property prediction (MPP) is a crucial task in drug discovery, used to predict the biological activity, toxicity, and pharmacokinetics of molecules through computational models, thereby accelerating the screening of candidate drugs. MPP methods mainly include sequence models based on the Simplified Linear Input Canonical Model (SMILES), fixed-length vector representation methods based on molecular fingerprints, and structured representation methods based on graph neural networks (GNNs). Among these, GNNs exhibit significant advantages due to their ability to directly model the graph structure of molecules (atoms as nodes, chemical bonds as edges). However, in practical applications, traditional supervised learning methods struggle to achieve ideal performance under small sample conditions due to high experimental validation costs and scarce labeled data.

[0003] To address the problem of data scarcity, few-shot learning techniques have been introduced into the field of molecular attribute prediction. Among them, meta-learning, through the "learning-to-learn" paradigm, simulates few-shot tasks during the training phase, enabling the model to quickly adapt to new tasks. For example, Meta-GNN combines the Model-Agnostic Meta-Learning (MAML) framework with GNNs for few-shot node classification tasks; GS-Meta (IJCAI) improves few-shot performance by constructing a molecular-attribute relationship graph and utilizing contrastive subgraph sampling. Furthermore, Attribute-Guided Prototype Networks (APNs) enhance generalization by introducing artificially defined molecular attributes. These methods perform well under few-shot conditions, but they neglect the complex relationships between tasks in molecular attribute prediction: molecules often possess multiple attributes simultaneously, and these attributes are dependent on each other (e.g., high lipophilicity versus low water solubility), while different attributes may be determined by different molecular substructures (e.g., benzene rings are effective for the NR-AiR task, while chloride ions are more critical for the SR-MMP task). This entangled relationship between tasks can lead to the model learning spurious or unstable features, affecting generalization performance.

[0004] In recent years, disentangled representation learning has shown promise in processing heterogeneous graph data, aiming to decompose complex relationships and representations into independent and semantically clear components. Compared to entangled representations, disentangled representations can reduce interference between irrelevant features and improve the robustness of models under few-shot conditions. In meta-learning, by separating task-invariant and task-specific features, shared knowledge across tasks and task-specific patterns can be captured more effectively. For example, the Global Imaging Principle (GIL) uses invariant learning to extract stable graph representations from different environments; the Integrated Generalized Model (IGM) mitigates spurious associations through subgraph fusion. However, existing methods have not fully explored the potential of task-to-task decoupling in molecular property prediction.

[0005] However, traditional methods suffer from problems such as inter-task entanglement, insufficient modeling of heterogeneous relationships, and sensitivity to environmental noise in predicting molecular attributes with few samples. Inter-task entanglement refers to the fact that existing meta-learning methods (such as MAML and GS-Meta) do not explicitly model the invariant and specific features between tasks, leading to the model potentially overfitting to spurious associations between tasks. For example, redundant dependencies between molecular attributes (such as the negative correlation between lipid solubility and water solubility) may mask key substructures unique to each task (such as the decisive role of oxide groups in the SR-MMP task). Insufficient modeling of heterogeneous relationships refers to the existence of complex many-to-many relationships between molecules and attributes (such as one molecule corresponding to multiple attributes, and one attribute associating with multiple molecules), which traditional graph structures (such as single-molecule graphs or attribute relationship graphs) cannot adequately capture. At the same time, existing methods lack robustness to noise in molecular data (such as experimental errors or missing labels) and have limited predictive ability for rare attributes (such as certain toxicity labels), mainly due to the problem of imbalanced samples. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a few-shot graph classification method and system based on graph decoupling meta-learning, in order to solve the technical problem of improving the accuracy and generalization ability of molecular attribute prediction.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution.

[0008] This invention first discloses a few-shot graph classification method based on graph decoupling meta-learning, which includes the following steps: Step 1: Construct a multi-relationship molecular graph and explicitly model molecular nodes, attribute nodes, and edges between nodes; Step 2: Construct a meta-learning task based on the multi-relation molecular graph, and use decoupling learning and meta-learning strategy models to probabilistically transform the substructure of the molecule, thereby decoupling the task-invariant features and task-specific features in the molecule; Step 3: Update the model parameters in the task-specific learning phase and the task-invariant learning phase respectively, and optimize the meta-learning process.

[0009] The present invention further includes the following preferred embodiments: The construction of the multi-relationship molecular graph, which explicitly models molecular nodes, attribute nodes, and edges between nodes, further includes: Through GNN encoder Extracting molecular nodes Features, randomly initialize attribute nodes Features:

[0010] Nodes in a multi-relationship molecular graph The initial feature vector; Indicates the first One molecule; A set of molecular nodes; A collection of attribute nodes; Molecular-attribute edge Direct association between known molecules and attribute tags, molecule-molecule edges Dynamically learn the similarity between molecules through attention weights Filtering Top-n links: ; in This refers to a multi-layer perceptron. For activation functions; Represents a node With nodes The distance between feature vectors.

[0011] The decoupling of task-invariant and task-specific features in the molecule further includes: Rationale features and Complement features were extracted using different encoders:

[0012] in, For specific molecules, Rationale represents the set of task-related features in a molecule, i.e., key substructures that are directly related to the target properties; The term "Complement" represents the set of task-independent features. Indicates the constituent units in a molecule; This is a preset proportional coefficient; Minimize the mutual information between the two by comparing the loss functions:

[0013] in The total number of tasks. Temperature parameters for adjusting the sharpness of the characteristic distribution, and Forming positive sample pairs and These represent cross-task negative samples and complement negative samples, respectively; positive sample pairs Strengthen the consistency of rationale features within the same task; double negative sample terms. Reasoning about rational features for other tasks, Then the complement feature is separated.

[0014] The step of updating model parameters in the task-specific learning phase and the task-invariant learning phase further includes: During the task-specific learning phase, keep the task-invariant parameters fixed and optimize the Rationale parameters:

[0015] in For Rationale related parameters, For learning rate, In response to gradient calculation, In the mission Support set With query set The classification loss is calculated above; where the classification loss is the cross entropy. During the task-invariant learning phase, the Rationale parameter is fixed, and the task-invariant parameter is updated through variance regularization: ; in For parameters related to Complement, For the corresponding learning rate, For variance regularization, Its regularization coefficient is denoted as .

[0016] This invention also discloses a few-shot graph classification system based on graph decoupling meta-learning, which utilizes the aforementioned few-shot graph classification method based on graph decoupling meta-learning, comprising: The molecular graph modeling module is used to construct multi-relationship molecular graphs and to explicitly model molecular nodes, attribute nodes, and edges between nodes. The decoupled meta-learning module is used to construct a meta-learning task based on the multi-relation molecular graph, and to probabilize the substructure of the molecule using the decoupled learning and meta-learning strategy model, thereby decoupling the task-invariant features and task-specific features in the molecule. The update and optimization module is used to update model parameters in both the task-specific learning phase and the task-invariant learning phase, thereby optimizing the meta-learning process.

[0017] Accordingly, this application also discloses a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the aforementioned few-shot graph classification method based on graph decoupling meta-learning.

[0018] Accordingly, this application also discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned few-shot graph classification method based on graph decoupling meta-learning.

[0019] The beneficial effects of this invention are that, compared with existing technologies, it provides a few-shot graph classification method and system based on graph decoupling meta-learning. By constructing multi-relational molecular graphs and employing innovative decoupling learning strategies, it demonstrates significant advantages in few-shot molecular attribute prediction tasks. By decoupling the task-specific and task-invariant parts of molecular features, the model not only improves prediction accuracy but also enhances feature interpretability. Visual analysis confirms a strong correlation between decoupling features and chemical properties, while the subgraph-based meta-learning framework effectively reduces computational complexity, enabling efficient processing of large-scale molecular databases. This technology provides a reliable solution to the few-shot attribute prediction problem in drug discovery and materials design. Attached Figure Description

[0020] Figure 1 This is a flowchart of the few-shot graph classification method based on graph decoupling meta-learning in this invention.

[0021] Figure 2 This is a schematic diagram comparing the classification results of the present invention and traditional methods on a benchmark dataset. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0023] The embodiments described in this application are merely some, not all, embodiments of the present invention. Based on the spirit of the present invention, other embodiments obtained by those skilled in the art without inventive effort are all within the protection scope of the present invention.

[0024] To address the shortcomings of existing technologies, this invention proposes a few-shot graph classification method and system based on graph decoupling meta-learning. It explicitly models the complex relationships between molecules and attributes by constructing a multi-relation molecular graph (MMG), and utilizes decoupling learning and meta-learning strategies to learn and separate task-invariant and task-specific features, achieving efficient feature extraction and knowledge transfer. First, task subgraphs are sampled from the MMG. Then, task-related "rationales" and their complements are generated through probabilistic molecular substructure extraction, capturing attribute-specific functional groups and cross-task shared molecular structures, respectively. Based on this, the meta-learning process is optimized by combining contrastive loss and invariant regularization, ensuring that the model can adapt to task-specific patterns while utilizing shared knowledge across tasks.

[0025] See Figure 1 As shown, the few-shot graph classification method based on graph decoupling meta-learning disclosed in this invention includes the following steps: Step 1: Construct a multi-relationship molecular graph and explicitly model molecular nodes, attribute nodes, and edges between nodes.

[0026] The many-to-many relationships between molecules and attributes are modeled using the Multi-Relationship Molecular Graph (MMG), and its structure is defined as follows: Graph node types include molecular nodes. and attribute nodes Molecular node features are processed by a GNN encoder. Extract and randomly initialize the features of the attribute nodes.

[0027]

[0028] in, Nodes in a multi-relationship molecular graph The initial feature vector; Indicates the first One molecule; A set of molecular nodes; This is a set of attribute nodes. When a node belongs to the set of sub-nodes... At that time, the initial features were generated by the GNN encoder. Extracted molecules Features are obtained when a node belongs to the set of attribute nodes. When initializing features, the types of edges randomly initialized include molecule-attribute edges ( ) and molecular-molecular edge ( Molecular-attribute edges refer to direct associations between known molecules and attribute labels. Molecular-molecular edges, on the other hand, refer to dynamically learning the similarity between molecules through attention weights. Filter the top-n connections.

[0029]

[0030] in This refers to a multi-layer perceptron. For activation functions; Represents a node With nodes The distance between feature vectors. The multilayer perceptron is used to perform a nonlinear mapping on the similarity between nodes, thereby calculating the molecule-molecule edge. Dynamic attention weights The construction of MMG allows for the explicit expression of complex dependencies between molecules and properties, providing a structured foundation for subsequent task decoupling.

[0031] Step 2: Construct a meta-learning task based on the multi-relation molecular graph, and use decoupling learning and meta-learning strategy models to probabilistically transform the substructure of the molecule, thereby decoupling the task-invariant features and task-specific features in the molecule.

[0032] For target attributes The meta-learning task is constructed by sampling local subgraphs from the MMG. The support set contains molecules and their labels directly associated with the target attribute, and the query set is used to evaluate model performance. Auxiliary attribute nodes and molecule similarity subgraphs are introduced to enhance task context information. Local subgraph sampling reduces computational complexity while preserving the global semantic associations between molecules and attributes.

[0033] Rational and complementary features in the molecule are separated through probabilistic modeling. Rational features are extracted based on GNN-based node importance prediction, sampling molecular substructures and their complements. Feature decoupling is achieved by extracting Rational and complementary features using different encoders.

[0034] in, For specific molecules, This represents the task-related feature set (Rationale) in a molecule, which is the key substructure directly related to the target attribute; This represents the task-independent feature set (Complement), which is the complement of the molecule; Indicates the constituent units (such as atoms or substructures) in a molecule; This is a preset scaling factor used to control the size of the task-related feature set. The mutual information between the two is minimized by comparing the loss function.

[0035] in The total number of tasks. Temperature parameters for adjusting the sharpness of the characteristic distribution, and Forming positive sample pairs and These represent cross-task negative samples and complement negative samples, respectively. Intramolecular positive sample pairs. To enhance the consistency of rationale features within the same task; secondly, a double negative sample term is introduced into the denominator, where... Reasoning about rational features for other tasks, Then, complementary features are separated; finally, cross-task stability is ensured through task averaging. Temperature parameter. By adjusting the sharpness of the feature distribution, explicit decoupling between task-specific features and shared features is ultimately achieved. This loss function allows Rationale to focus on task-specific patterns (such as functional groups), while Complement preserves cross-task shared structures (such as benzene rings).

[0036] Step 3: Update the model parameters in the task-specific learning phase and the task-invariant learning phase respectively, and optimize the meta-learning process.

[0037] This implementation combines a MAML framework to achieve task adaptation and knowledge transfer. During the task-specific learning phase, the task-specific parameters are kept constant while the Rationale parameters are optimized.

[0038] in For Rationale related parameters, For learning rate, In response to gradient calculation, In the mission Support set With query set The classification loss is calculated above.

[0039] During the task-invariant learning phase, the Rationale parameter is fixed, and the task-invariant parameter is updated through variance regularization:

[0040] in For parameters related to task invariance (Complement), For the corresponding learning rate, This is a variance regularization term used to constrain loss fluctuations across different tasks to improve generalization. This is its regularization coefficient. It constrains the loss fluctuations across different tasks to improve generalization.

[0041] See Figure 2 Experiments show that the present invention significantly outperforms existing methods on five benchmark datasets, including Tox21 and SIDER (e.g., a 6.10% improvement in 1-shot performance on the PCBA dataset), and the effectiveness of the decoupled representation is verified through visualization.

[0042] The beneficial effects of this invention are that, compared with existing technologies, it provides a few-shot graph classification method and system based on graph decoupling meta-learning. By constructing multi-relational molecular graphs and employing innovative decoupling learning strategies, it demonstrates significant advantages in few-shot molecular attribute prediction tasks. By decoupling the task-specific and task-invariant parts of molecular features, the model not only improves prediction accuracy but also enhances feature interpretability. Visual analysis confirms a strong correlation between decoupling features and chemical properties, while the subgraph-based meta-learning framework effectively reduces computational complexity, enabling efficient processing of large-scale molecular databases. This technology provides a reliable solution to the few-shot attribute prediction problem in drug discovery and materials design.

[0043] This invention can be a system, method, and / or computer program product. This invention also discloses a few-shot graph classification system based on graph decoupling meta-learning, which is based on the aforementioned few-shot graph classification method based on graph decoupling meta-learning, comprising: The molecular graph modeling module is used to construct multi-relationship molecular graphs and to explicitly model molecular nodes, attribute nodes, and edges between nodes. The decoupled meta-learning module is used to construct a meta-learning task based on the multi-relation molecular graph, and to probabilize the substructure of the molecule using the decoupled learning and meta-learning strategy model, thereby decoupling the task-invariant features and task-specific features in the molecule. The update and optimization module is used to update model parameters in both the task-specific learning phase and the task-invariant learning phase, thereby optimizing the meta-learning process.

[0044] Based on the spirit of this invention, those skilled in the art will readily conceive of a computer program product derived from the aforementioned few-shot graph classification method based on graph decoupling meta-learning. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure. That is, this application also includes a terminal comprising a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to perform the steps of the aforementioned few-shot graph classification method based on graph decoupling meta-learning.

[0045] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example, but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0046] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0047] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0048] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A few-shot graph classification method based on graph decoupling meta-learning, characterized in that, Includes the following steps: Step 1: Construct a multi-relationship molecular graph and explicitly model molecular nodes, attribute nodes, and edges between nodes; Step 2: Construct a meta-learning task based on the multi-relation molecular graph, and use decoupling learning and meta-learning strategy models to probabilistically transform the substructure of the molecule, thereby decoupling the task-invariant features and task-specific features in the molecule; Step 3: Update the model parameters in the task-specific learning phase and the task-invariant learning phase respectively, and optimize the meta-learning process.

2. The few-shot graph classification method based on graph decoupling meta-learning according to claim 1, characterized in that, The construction of the multi-relationship molecular graph, which explicitly models molecular nodes, attribute nodes, and edges between nodes, further includes: Through GNN encoder Extracting molecular nodes Features, randomly initialize attribute nodes Features: Nodes in a multi-relationship molecular graph The initial feature vector; Indicates the first One molecule; A set of molecular nodes; A collection of attribute nodes; Molecular-attribute edge Direct association between known molecules and attribute tags, molecule-molecule edges Dynamically learn the similarity between molecules through attention weights Filtering Top-n links: ； in This refers to a multi-layer perceptron. For activation functions; Represents a node With nodes The distance between feature vectors.

3. The few-shot graph classification method based on graph decoupling meta-learning according to claim 2, characterized in that, The decoupling of task-invariant and task-specific features in the molecule further includes: Rationale features and Complement features were extracted using different encoders: in, For specific molecules, Rationale represents the set of task-related features in a molecule, i.e., key substructures that are directly related to the target properties; The term "Complement" represents the set of task-independent features. Indicates the constituent units in a molecule; This is a preset proportional coefficient; Minimize the mutual information between the two by comparing the loss functions: in The total number of tasks. Temperature parameters for adjusting the sharpness of the characteristic distribution, and Forming positive sample pairs and These represent cross-task negative samples and complement negative samples, respectively; positive sample pairs Strengthen the consistency of rationale features within the same task; double negative sample terms. Reasoning about rational features for other tasks, Then the complement feature is separated.

4. The few-shot graph classification method based on graph decoupling meta-learning according to claim 3, characterized in that, The step of updating model parameters in the task-specific learning phase and the task-invariant learning phase further includes: During the task-specific learning phase, keep the task-invariant parameters fixed and optimize the Rationale parameters: in For Rationale related parameters, For learning rate, In response to gradient calculation, In the mission Support set With query set The classification loss is calculated above; where the classification loss is the cross entropy. During the task-invariant learning phase, the Rationale parameter is fixed, and the task-invariant parameter is updated through variance regularization: ； in For parameters related to Complement, For the corresponding learning rate, For variance regularization, Its regularization coefficient is denoted as .

5. A few-shot graph classification system based on graph decoupling meta-learning, characterized in that, include: The molecular graph modeling module is used to construct multi-relationship molecular graphs and to explicitly model molecular nodes, attribute nodes, and edges between nodes. The decoupled meta-learning module is used to construct a meta-learning task based on the multi-relation molecular graph, and to probabilize the substructure of the molecule using the decoupled learning and meta-learning strategy model, thereby decoupling the task-invariant features and task-specific features in the molecule. The update and optimization module is used to update model parameters in both the task-specific learning phase and the task-invariant learning phase, thereby optimizing the meta-learning process.

6. The few-shot graph classification system based on graph decoupling meta-learning according to claim 5, characterized in that, The molecular graph modeling module is further used for: Through GNN encoder Extracting molecular nodes Features, randomly initialize attribute nodes Features: Nodes in a multi-relationship molecular graph The initial feature vector; Indicates the first One molecule; A set of molecular nodes; A collection of attribute nodes; Molecular-attribute edge Direct association between known molecules and attribute tags, molecule-molecule edges Dynamically learn the similarity between molecules through attention weights Filtering Top-n links: ； in This refers to a multi-layer perceptron. For activation functions; Represents a node With nodes The distance between feature vectors.

7. The few-shot graph classification system based on graph decoupling meta-learning according to claim 6, characterized in that, The decoupling meta-learning module is further used for: Rationale features and Complement features were extracted using different encoders: in, For specific molecules, Rationale represents the set of task-related features in a molecule, i.e., key substructures that are directly related to the target properties; The term "Complement" represents the set of task-independent features. Indicates the constituent units in a molecule; This is a preset proportional coefficient; Minimize the mutual information between the two by comparing the loss functions: in The total number of tasks. Temperature parameters for adjusting the sharpness of the characteristic distribution, and Forming positive sample pairs and These represent cross-task negative samples and complement negative samples, respectively; positive sample pairs Strengthen the consistency of rationale features within the same task; double negative sample terms. Reasoning about rational features for other tasks, Then the complement feature is separated.

8. The few-shot graph classification system based on graph decoupling meta-learning according to claim 7, characterized in that, The update and optimization module is further used for: During the task-specific learning phase, keep the task-invariant parameters fixed and optimize the Rationale parameters: in For Rationale related parameters, For learning rate, In response to gradient calculation, In the mission Support set With query set The classification loss is calculated above; where the classification loss is the cross entropy. During the task-invariant learning phase, the Rationale parameter is fixed, and the task-invariant parameter is updated through variance regularization: in For parameters related to Complement, For the corresponding learning rate, For variance regularization, Its regularization coefficient is denoted as .

9. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the few-shot graph classification method based on graph decoupling meta-learning according to any one of claims 1-4.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the few-shot graph classification method based on graph decoupling meta-learning as described in any one of claims 1-4.

Images