Decoupling representation learning-based out-of-distribution generalization graph processing method, system and product

Abstract

The application provides a distribution out-of-distribution generalization graph processing method and system based on decoupled representation learning and a product, and relates to the technical field of graph processing. The method comprises the following steps: inputting unlabeled graph structure data into a graph neural network model to be trained, performing L times of iterative updates on the node representation of each node in the graph structure data according to neighbor information to obtain updated node representation; processing the updated node representation through K graph convolution channels to obtain K representations of each node; processing the K representations of each node into a representation vector of graph-invariant information and a representation vector of graph-variant information; simulating multiple distribution environments according to the representation vector of graph-variant information; under the multiple distribution environments, performing self-supervised contrast learning on the representation vector of graph-invariant information to obtain a trained graph neural network model and use the trained graph neural network model for classification of a graph dataset, so as to adaptively ensure the prediction effect of the graph neural network model when there is a distribution difference between a test environment and a training environment.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of graph processing technology, and in particular to a distributed out-generalization graph processing method, system and product based on decoupled representation learning. Background Technology

[0002] In the real world, graph-structured data is ubiquitous across various fields, such as social networks, transportation networks, financial networks, and chemical molecules. Graph Neural Networks (GNNs), as a core topic in graph machine learning, have made significant progress in both academia and industry.

[0003] However, most existing studies assume that training and testing data come from the same distribution, ignoring the situation where the distributions of training and testing data are inconsistent in real-world applications. This distribution shift is prevalent and unavoidable, leading to a significant performance degradation of existing graph neural networks when faced with distribution shifts. Currently, existing work heavily relies on graph labels when learning out-of-distribution generalization graph representations. However, graph labels are often scarce or even unavailable in real-world applications. This results in low performance of graph neural networks and limited graph representation learning ability in out-of-distribution generalization scenarios. Summary of the Invention

[0004] To address the aforementioned technical issues, this invention provides a method, system, and product for processing out-of-distribution generalization graphs based on decoupled representation learning, which adaptively ensures the prediction performance of graph neural network models when there are distribution differences between the testing and training environments.

[0005] This invention provides a method for processing out-of-distribution generalization graphs based on decoupled representation learning, the method comprising:

[0006] Unlabeled graph structure data is input into the graph neural network model to be trained. The neighbor aggregation mechanism is used to update the node representation of each node in the unlabeled graph structure data L times according to the neighbor information to obtain the updated node representation. The updated node representation of a node aggregates all the information of the node's L-order neighbors.

[0007] The updated node representations are processed through the K independent graph convolution channels of the graph neural network model to be trained to obtain K representations for each node; where K and L are natural numbers greater than 1.

[0008] The K representations of each node are processed into a representation vector of graph-invariant information and a representation vector of graph-change information;

[0009] The graph neural network model to be trained simulates various distribution environments of the unlabeled graph structure data based on the representation vector of the graph change information.

[0010] Using the graph neural network model to be trained, in the various distribution environments, self-supervised comparative learning is performed using the representation vector of the graph invariant information to obtain a trained graph neural network model.

[0011] The molecular graphs in the graph dataset are input into the trained graph neural network model to obtain the graph representation of the molecular graphs, and the graph dataset is classified based on the graph representation of the molecular graphs.

[0012] Optionally, processing the K representations of each node into a representation vector of graph-invariant information and a representation vector of graph-change information includes:

[0013] Based on the idea of ​​group theory, the graph information classification neural network assumes that any one or more of the K representations are the representation vectors of candidate graph change information, and assumes that the remaining representations of the K representations are the representation vectors of candidate graph invariant information.

[0014] Based on the representation vector of the candidate graph change information, various hypothetical distribution environments are simulated;

[0015] Under the various hypothetical distribution environments, self-supervised comparative learning is performed using the representation vector of the candidate graph invariant information to calculate the hypothesis loss value corresponding to the graph information classification neural network.

[0016] Re-assume that any one or more of the K representations are representation vectors of the candidate graph change information, and assume that the remaining representations of the K representations are representation vectors of the candidate graph invariant information, until the assumption loss values ​​of all assumption cases of the K representations are calculated;

[0017] Based on the assumptions of the K representations corresponding to the minimum hypothesis loss value, the K representations are processed into representation vectors of graph invariant information and representation vectors of graph change information.

[0018] Optionally, simulating various distribution environments of the unlabeled graph structure data based on the representation vector of the graph change information includes:

[0019] Clustering is performed based on the representation vectors of the graph change information to obtain the clustering results;

[0020] Based on the clustering results, the various distribution environments are simulated to encourage the representation vectors of graph change information to maintain predictability of the distribution environment in the various distribution environments.

[0021] Optionally, the graph neural network model to be trained includes at least: a normalizer; and under the various distribution environments, self-supervised contrastive learning is performed using the representation vectors of the graph-invariant information to obtain a trained graph neural network model, including:

[0022] The normalizer enables the graph-invariant information representation vector to maintain stability in graph representation prediction across various distribution environments. The graph-invariant information representation vector represents the information in the unlabeled graph structure data that truly has graph representation prediction value.

[0023] Graph representation prediction is performed based on the representation vector of the graph invariant information, and the loss value corresponding to the graph neural network model to be trained is calculated through self-supervised contrastive learning.

[0024] Based on the hypothetical loss value corresponding to the graph information classification neural network and the loss value corresponding to the graph neural network model to be trained, the model parameters of the graph neural network model to be trained are adjusted to obtain the trained graph neural network model.

[0025] Optionally, the graph dataset includes a drug dataset, the molecular graph includes a drug molecule graph, and the method further includes:

[0026] The trained graph neural network model was tested using labeled molecular graphs to obtain test results;

[0027] The step of inputting molecular graphs from the graph dataset into the trained graph neural network model to obtain graph representations of the molecular graphs, and classifying the graph dataset based on the graph representations of the molecular graphs, includes:

[0028] If the test results meet the conditions, the unlabeled drug molecule graphs in the drug dataset are input into the trained graph neural network model to obtain the graph representation of the drug molecule graphs.

[0029] The drug dataset is classified based on the graph representation of the drug molecule graph.

[0030] Optionally, the method further includes:

[0031] The social network graph data is input into the trained graph neural network model to obtain the graph representation of the social network graph data.

[0032] Based on the graph representation of the social network graph data, the social network graph data is analyzed to obtain analysis results;

[0033] The analysis results include at least the dynamic changes and data migration of the social networks corresponding to the social network graph data.

[0034] Optionally, the method further includes:

[0035] The graph data in the human-computer interaction system is input into the trained graph neural network model to obtain the graph representation corresponding to the human-computer interaction.

[0036] Based on the graph representation corresponding to the human-computer interaction, user needs and user behaviors in the human-computer interaction system are understood and predicted.

[0037] A second aspect of this invention provides a distributed out-generalization graph processing system based on decoupled representation learning, the system comprising:

[0038] The node representation module is used to input unlabeled graph structure data into the graph neural network model to be trained. It adopts a neighbor aggregation mechanism to update the node representation of each node in the unlabeled graph structure data through L iterations based on the neighbor information to obtain the updated node representation. The updated node representation of a node aggregates all the information of the node's L-order neighbors.

[0039] The representation processing module is used to process the updated node representations through K independent graph convolution channels of the graph neural network model to be trained, so as to obtain K representations for each node; where K and L are natural numbers greater than 1.

[0040] The representation classification module is used to process the K representations of each node into a representation vector of graph-invariant information and a representation vector of graph-change information;

[0041] The environment simulation module is used to simulate various distribution environments of the unlabeled graph structure data based on the representation vector of the graph change information through the graph neural network model to be trained.

[0042] The model training module is used to perform self-supervised comparative learning using the graph invariant information representation vectors of the graph neural network model to be trained under the various distribution environments, so as to obtain a trained graph neural network model.

[0043] The graph processing module is used to input the molecular graphs in the graph dataset into the trained graph neural network model to obtain the graph representation of the molecular graphs, and to classify the graph dataset based on the graph representation of the molecular graphs.

[0044] Optionally, the characterization classification module includes:

[0045] The first assumption module is used to classify neural networks through graph information. Based on the idea of ​​group theory, it assumes that any one or more of the K representations are representation vectors of candidate graph change information, and assumes that the remaining representations of the K representations are representation vectors of candidate graph invariant information.

[0046] The environment hypothesis module is used to simulate multiple hypothesis distribution environments based on the representation vector of the candidate graph change information;

[0047] The first loss calculation module is used to perform self-supervised comparative learning using the representation vector of the candidate graph invariant information under the multiple hypothesis distribution environments, and calculate the hypothesis loss value corresponding to the graph information classification neural network.

[0048] The second hypothesis module is used to re-hypothesize any one or more of the K representations as the representation vector of the candidate graph change information, and to assume that the remaining representations of the K representations are the representation vector of the candidate graph invariant information, until the hypothesis loss value of all hypothetical cases of the K representations is calculated.

[0049] The representation determination module is used to process the K representations corresponding to the minimum hypothesis loss value into a representation vector of the graph invariant information and a representation vector of the graph change information, based on the hypothesis of the K representations.

[0050] Optionally, the environment simulation module includes:

[0051] The clustering module is used to perform clustering based on the representation vector of the graph change information to obtain clustering results;

[0052] An environment determination module is used to simulate the various distribution environments based on the clustering results, so as to encourage the representation vector of the graph change information to maintain predictability of the distribution environment in the various distribution environments.

[0053] Optionally, the graph neural network model to be trained includes at least: a regularizer; and a model training module, including:

[0054] The normalization module is used to ensure that the representation vector of the graph invariant information maintains stability in graph representation prediction in various distribution environments through the normalizer. The representation vector of the graph invariant information represents the information in the unlabeled graph structure data that truly has graph representation prediction value.

[0055] The second loss calculation module is used to perform graph representation prediction based on the representation vector of the graph invariant information, and calculate the loss value corresponding to the graph neural network model to be trained through self-supervised contrastive learning.

[0056] The model parameter tuning module is used to adjust the model parameters of the graph neural network model to be trained based on the hypothesis loss value corresponding to the graph information classification neural network and the loss value corresponding to the graph neural network model to be trained, so as to obtain the trained graph neural network model.

[0057] Optionally, the graph dataset includes a drug dataset, the molecular graph includes a drug molecule graph, and the system further includes:

[0058] The model testing module is used to test the trained graph neural network model using labeled molecular graphs and obtain test results.

[0059] The graph processing module includes:

[0060] The drug characterization module is used to input the unlabeled drug molecule graphs in the drug dataset into the trained graph neural network model to obtain the graph representation of the drug molecule graphs when the test results meet the conditions.

[0061] The drug classification module is used to classify the drug dataset based on the graph representation of the drug molecule graph.

[0062] Optionally, the system further includes:

[0063] The social data processing module is used to input social network graph data into the trained graph neural network model to obtain a graph representation of the social network graph data.

[0064] The social data analysis module is used to analyze the social network graph data based on the graph representation of the social network graph data and obtain analysis results.

[0065] The analysis results include at least the dynamic changes and data migration of the social networks corresponding to the social network graph data.

[0066] Optionally, the system further includes:

[0067] The human-computer interaction processing module is used to input the graph data of the human-computer interaction system into the trained graph neural network model to obtain the graph representation corresponding to the human-computer interaction.

[0068] The human-computer interaction analysis module is used to understand and predict user needs and user behaviors in the human-computer interaction system based on the graph representation corresponding to the human-computer interaction.

[0069] A third aspect of the present invention provides an electronic device, the electronic device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, it implements the distribution out-generalization graph processing method based on decoupled representation learning as described in the first aspect of the present invention.

[0070] A fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the distribution out-generalization graph processing method based on decoupled representation learning as described in the first aspect of the present invention.

[0071] This invention provides a method for out-of-distribution generalization graph processing based on decoupled representation learning. First, unlabeled graph structure data is input into a graph neural network model to be trained. A neighbor aggregation mechanism is used to iteratively update the node representation of each node in the unlabeled graph structure data L times, resulting in updated node representations. Second, the updated node representations of each node are processed through K independent graph convolution channels, resulting in K representations for each node. Then, the K representations of each node are processed into graph-invariant information representation vectors and graph-change information representation vectors. Based on the graph-change information representation vectors, various distribution environments of the unlabeled graph structure data are simulated. Finally, under various distribution environments, self-supervised comparative learning is performed using the graph-invariant information representation vectors to obtain... A trained graph neural network (GNN) model is used to classify molecular graphs. This allows for the differentiation, independence, and dissociation of graph-invariant and graph-variant information in the graph structure representation without relying on data labels, thereby removing spurious correlations in the representation. Multiple distribution environments are simulated based on graph-variant information, and self-supervised comparative learning is performed in these environments based on graph-invariant information. This prevents the GNN model from making predictions based on spurious correlations between graph-variant information and labels, encouraging it to make predictions based on stable graph-invariant information. This adaptively ensures the model's prediction performance even when there are distribution differences between the test and training environments, improving the prediction performance of out-of-distribution generalization GNNs and enhancing the graph representation learning ability of GNNs. Attached Figure Description

[0072] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] Figure 1 This is a flowchart illustrating a distributional out-generalization graph processing method based on decoupled representation learning, provided in an embodiment of the present invention.

[0074] Figure 2 This is a flowchart illustrating a distributional out-generalization graph pre-training method based on decoupled representation, as shown in an embodiment of the present invention.

[0075] Figure 3This is a structural block diagram of a distributed out-generalization graph processing system based on decoupled representation learning, as shown in an embodiment of the present invention.

[0076] Figure 4 This is a schematic diagram of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0077] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0078] Currently, graph neural networks are mainly divided into three aspects:

[0079] The first aspect is graph neural network representation learning: Graph neural networks have recently gained widespread attention due to their powerful graph representation capabilities, achieving relatively ideal performance on many tasks. These methods primarily employ a neighbor aggregation mechanism, iteratively updating node representations based on neighbor information. After obtaining the updated node representations, graph pooling is used to aggregate and summarize the entire graph representation. However, this type of method suffers from reduced performance when the distributions of the testing and training environments are inconsistent, thus limiting its application scenarios and making it unsuitable for real-world scenarios involving out-of-distribution generalization. Out-of-distribution generalization refers to situations where the data distribution used during model training differs from the data distribution encountered in actual applications.

[0080] The second aspect is out-of-distribution generalization graph neural networks: these methods do not depend on the number of nodes in the graph structure data itself to provide prediction results, and can transfer models trained on small graphs to large graphs, achieving good results on large graphs. However, this type of method only adapts to changes in the number of nodes in the training and testing environments, and cannot adapt to other types of distribution changes. Similarly, it cannot be used for out-of-distribution generalization graph representation learning problems in real-world scenarios.

[0081] The third aspect is self-supervised pre-trained graph neural networks: Self-supervised pre-training is a popular unsupervised method in computer vision and natural language processing. It has revolutionized graph learning due to its excellent performance in learning representations from cleverly designed prior tasks without relying on labels. Self-supervised pre-trained graph neural networks are mainly divided into two categories: contrastive and generative methods. Contrastive pre-trained graph neural networks use instance discrimination as the pre-training task, while generative pre-trained graph neural networks use direct reconstruction of missing parts of the input graph data as the pre-training task. However, these methods cannot be applied to situations where the training and testing environments are inconsistent. Therefore, the obtained representations cannot accurately depict the graph structure data under unknown testing conditions, which to some extent affects the performance of downstream tasks.

[0082] It is evident that current graph neural networks (GNNs) only fit the training environment to their vectorized representations, failing to adapt to unknown, complex, and potentially changing test environments. They primarily focus on fitting the training data, representing a holistic vector from a global perspective, thus failing to capture the crucial generalization information behind the graph structure data and neglecting to consider the true characteristics of the graph structure data when labels are missing. However, a large amount of unlabeled graph structure data in the real world often originates from complex distributions. When the training data is insufficient to reflect the true distribution, the test data will exhibit distributional differences from the training data. Therefore, in such cases, traditional methods may achieve good results on the training dataset and terminate training, but when applied to the test environment, performance will significantly degrade due to distribution shift, greatly limiting application scenarios and preventing the improvement of the graph neural network's representation learning ability through pre-training.

[0083] Therefore, in order to at least partially solve one or more of the above-mentioned problems and other potential problems, this invention proposes an out-of-distribution generalization graph processing method based on decoupled representation learning. In this method, without relying on data labels, the graph-invariant information and graph-change information represented by nodes in the graph data structure are distinguished, independent, and decoupled. False correlations in the representations are removed, preventing the model from making predictions based on the false correlation between graph-change information and labels. Graph-change information is used to simulate multiple distribution environments, and graph-invariant information is used to make predictions in multiple distribution environments. This adaptively ensures the prediction performance of the graph neural network model when there are distribution differences between the test environment and the training environment, improves the prediction performance of the out-of-distribution generalization graph neural network, and enhances the graph representation learning ability of the graph neural network.

[0084] Reference Figure 1 , Figure 1 This is a flowchart illustrating a distributional out-of-generalization graph processing method based on decoupled representation learning, provided in an embodiment of the present invention. Figure 1 As shown, this method may include the following steps:

[0085] Step S11: Input the unlabeled graph structure data into the graph neural network model to be trained, and use the neighbor aggregation mechanism to update the node representation of each node in the unlabeled graph structure data L times according to the neighbor information to obtain the updated node representation; wherein, the updated node representation of a node aggregates all the information of the node's L-order neighbors.

[0086] In this embodiment, an unlabeled graph data structure is used as training samples to train the graph neural network model, enabling the trained model to perform graph representation prediction in out-of-distribution generalization scenarios. The unlabeled graph data structure refers to any graph data structure without labels, such as molecular graphs, social network graphs, transportation network graphs, financial network graphs, etc. This embodiment does not impose any limitations on this.

[0087] After inputting unlabeled graph structure data into the graph neural network model to be trained, a neighbor aggregation mechanism can be used to iteratively update the node representation of each node in the unlabeled graph structure data L times based on the neighbor information, resulting in the updated node representation. Neighbor aggregation is performed on the node information of each input unlabeled graph structure data, using the neighbor information of each node in the unlabeled graph structure data to update the information of each node (i.e., the current node), resulting in the updated node representation. Through an independent graph convolution channel, L iterative updates are performed, aggregating all information from each node's L-order neighbors, resulting in the updated node representation of each node. Specifically, the updated node representation of a node aggregates all information from that node's L-order neighbors.

[0088] Step S12: Process the updated node representations through the K independent graph convolution channels of the graph neural network model to be trained to obtain K representations for each node; where K and L are natural numbers greater than 1.

[0089] In this embodiment, after obtaining the updated node representation of each node, the updated node representation of each node can be processed through the K independent graph convolution channels of the graph neural network model to be trained, and the relationships of surrounding neighboring nodes can be aggregated independently to achieve decoupling of the hidden factors behind the unlabeled graph structure data, thus obtaining K representations of each node. In this way, the node representation of each node in the graph structure data is processed into K independent representations.

[0090] In this embodiment, K and L are natural numbers greater than 1, and K is a predefined model hyperparameter that can be arbitrarily set manually based on experience. This embodiment does not limit the specific values ​​of K and L.

[0091] Step S13: Process the K representations of each node into a representation vector of graph-invariant information and a representation vector of graph-change information.

[0092] In this embodiment, since the obtained vector representations (i.e., the K representations of each node) contain both graph-invariant information (such as stability information) and graph-change information (such as noise information) from the unlabeled graph structure data, the interrelationships between these information may degrade the performance of the graph neural network model when the training and testing environments differ significantly. Therefore, after obtaining the K representations of each node in the unlabeled graph structure data, this embodiment can process the K representations of each node into: a representation vector for graph-invariant information and a representation vector for graph-change information. For example, by converting the graph-invariant information and graph-change information identified from the K representations of each node into corresponding representation vectors, we obtain the representation vectors for graph-invariant information and graph-change information.

[0093] Step S14: Using the graph neural network model to be trained, simulate various distribution environments of the unlabeled graph structure data based on the representation vector of the graph change information.

[0094] In this embodiment, the graph neural network model to be trained can simulate various distribution environments of unlabeled graph structure data based on the representation vector of the obtained graph change information.

[0095] Step S15: Using the graph neural network model to be trained, self-supervised comparative learning is performed under the various distribution environments using the representation vector of the graph invariant information to obtain the trained graph neural network model.

[0096] In this embodiment, after simulating multiple distribution environments, the graph neural network model to be trained can be trained by using the graph invariant information representation vectors obtained in these multiple distribution environments to perform self-supervised comparative learning, thereby obtaining a trained graph neural network model.

[0097] For example, in an optional example, unlabeled graph structure data can be input into the graph neural network model to be trained in batches during each training process to obtain prediction results. Then, through self-supervised contrastive learning, the loss function is calculated, the gradient of the graph neural network model parameters is calculated based on the loss value corresponding to the obtained prediction results, and gradient updates are performed to optimize the performance of the graph neural network model until the training is completed, and a trained graph neural network model is obtained.

[0098] Step S16: Input the molecular graphs in the graph dataset into the trained graph neural network model to obtain the graph representation of the molecular graphs, and classify the graph dataset according to the graph representation of the molecular graphs.

[0099] In this embodiment, after obtaining the trained graph neural network, the molecular graphs in the graph dataset can be input into the trained graph neural network model for graph representation prediction. This yields the graph representations of the molecular graphs output by the trained graph neural network model, which are then used to classify the molecular graphs in the graph dataset, thus achieving graph dataset classification. The trained graph neural network model in this embodiment can be used to perform at least one or more of the following tasks: classification tasks and regression tasks.

[0100] In this embodiment, without relying on data labels, the graph-invariant information and graph-variable information represented by graph structure data are distinguished, independent, and decoupled to remove spurious correlations in the representation. Based on the graph-variable information, various distribution environments are simulated, and then self-supervised comparative learning is performed based on the graph-invariant information in various distribution environments. This prevents the graph neural network model from making predictions based on spurious correlations between graph-variable information and labels, and encourages the graph neural network model to make predictions based on stable graph-invariant information. This adaptively ensures the model's prediction performance when there are distribution differences between the test environment and the training environment, improves the prediction performance of the out-of-distribution generalization graph neural network, and enhances the graph representation learning ability of the graph neural network.

[0101] In other words, the method in this embodiment enables the trained graph neural network model to adaptively generate stable graph representations on out-of-distribution graph structure data, effectively improving the accuracy and stability of graph structure data processing and representation learning, and enhancing the adaptability and stability of the graph neural network model in test environments with different distributions. It does not rely on label information and can still work effectively even when data labels are missing, expanding the application scenarios and making it suitable for various practical applications, such as social network analysis, bioinformatics, and complex network system management. At the same time, by removing noise and spurious correlations (graph change information), the generalization ability of the model is further enhanced.

[0102] In conjunction with the above embodiments, in one implementation, the present invention also provides a method for processing out-of-distribution generalization graphs based on decoupled representation learning. In this method, step S13 may specifically include steps S21 to S25:

[0103] Step S21: Using a graph information classification neural network, based on the idea of ​​group theory, assume that any one or more of the K representations are the representation vectors of candidate graph change information, and assume that the remaining representations of the K representations are the representation vectors of candidate graph invariant information.

[0104] In this embodiment, the graph information classification neural network is a graph neural network used to process the K representations of each node into representation vectors of graph-invariant information and representation vectors of graph-change information. Based on group theory, this graph information classification neural network can assume that any one or more of the K representations are candidate representation vectors of graph-change information, and that the remaining K representations are candidate representation vectors of graph-invariant information. This is one possible assumption regarding the K representations.

[0105] Among them, the representation vector of candidate graph change information is the representation vector of assumed graph change information, and the representation vector of candidate graph invariance information is the representation vector of assumed graph invariance information.

[0106] Step S22: Based on the representation vector of the candidate graph change information, simulate multiple hypothetical distribution environments.

[0107] In this embodiment, after obtaining a hypothetical scenario, the representation vector of the hypothesized candidate graph change information is used to simulate various hypothetical distribution environments of the unlabeled graph structure data. The hypothetical distribution environment is the distribution environment simulated using the representation vector of the hypothesized graph change information.

[0108] Step S23: Under the multiple hypothesis distribution environment, self-supervised comparative learning is performed using the representation vector of the candidate graph invariant information to calculate the hypothesis loss value corresponding to the graph information classification neural network.

[0109] In this embodiment, self-supervised comparative learning can be performed using the representation vectors of candidate graph invariant information under various hypothetical distribution environments to calculate the hypothesis loss value corresponding to the graph information classification neural network. The hypothesis loss value is the loss value of the loss function corresponding to the graph information classification neural network.

[0110] Step S24: Re-assume that any one or more of the K representations are the representation vectors of the candidate graph change information, and assume that the remaining representations of the K representations are the representation vectors of the candidate graph invariant information, until the assumption loss values ​​of all the assumption cases of the K representations are calculated.

[0111] In this embodiment, the graph information classification neural network can assume all possible hypotheses for the K representations based on group theory and calculate the corresponding hypothesis loss value. That is, after assuming one hypothesis for the K representations and calculating the hypothesis loss value for that hypothesis, the graph information classification neural network will re-assume another hypothesis for the K representations: re-assume one or more of the K representations as representation vectors of candidate graph change information, re-assume the remaining representations in the current K representations as representation vectors of candidate graph invariant information, and calculate the hypothesis loss value for this other hypothesis. This process continues (e.g., for each hypothesis, steps S21 to S23 are re-executed) until all possible hypotheses for the K representations are assumed, and the hypothesis loss value for all possible hypotheses for the K representations is obtained.

[0112] For example, K is 3, and each node has 3 representations. First, we can assume that the first representation is the representation vector of candidate graph change information, and the second and third representations are the representation vectors of candidate graph invariant information. We use the first representation to simulate various hypothesis distribution environments, and the second and third representations are used for self-supervised comparative learning under these environments to calculate the corresponding hypothesis loss. Second, we can assume that the second representation is the representation vector of candidate graph change information, and the first and third representations are the representation vectors of candidate graph invariant information. We use the second representation to simulate various hypothesis distribution environments, and the first and third representations are used for self-supervised comparative learning under these environments to calculate the corresponding hypothesis loss. Then, we can assume that the third representation is the representation vector of candidate graph change information, and the first and second representations are the representation vectors of candidate graph invariant information. We use the third representation to simulate various hypothesis distribution environments, and the first and second representations are used for self-supervised comparative learning under these environments to calculate the corresponding hypothesis loss. We can assume that the first and second representations of the three representations are representation vectors of candidate graph change information, and the third representation is a representation vector of candidate graph invariant information. We can use the first and second representations to simulate various hypothesis distribution environments, and use the third representation for self-supervised comparative learning under these environments to calculate the corresponding hypothesis loss value. Similarly, we can assume that the second and third representations of the three representations are representation vectors of candidate graph change information, and the first representation is a representation vector of candidate graph invariant information. We can use the second and third representations to simulate various hypothesis distribution environments, and use the first representation for self-supervised comparative learning under these environments to calculate the corresponding hypothesis loss value. Finally, we can assume that the first and third representations of the three representations are representation vectors of candidate graph change information, and the second representation is a representation vector of candidate graph invariant information. We can use the first and third representations to simulate various hypothesis distribution environments, and use the second representation for self-supervised comparative learning under these environments to calculate the corresponding hypothesis loss value. This yields the hypothesis loss values ​​for all hypothetical scenarios of the three representations.

[0113] Step S25: Based on the assumptions of the K representations corresponding to the minimum hypothesis loss value, process the K representations into a representation vector of the graph invariant information and a representation vector of the graph change information.

[0114] In this embodiment, after obtaining the hypothesis loss values ​​for all hypothetical scenarios, the minimum hypothesis loss value can be determined among all the hypothesis loss values. The hypothesis scenario corresponding to the K representations of the minimum hypothesis loss value is the best-performing hypothesis scenario. The best-performing hypothesis scenario includes a representation vector of the hypothetical graph-invariant information and a representation vector of the hypothetical graph-change information. Therefore, based on the hypothesis scenario of the K representations corresponding to the minimum hypothesis loss value, the K representations are processed into representation vectors of graph-invariant information and representation vectors of graph-change information.

[0115] For example, continuing with the above three representations, if the assumptions of the K representations corresponding to the minimum hypothesis loss value are as follows: assuming that the first and third representations of the three representations are representation vectors of candidate graph change information, and assuming that the second representation of the three representations is representation vector of candidate graph invariant information, then the graph information classification neural network will ultimately determine the first and third representations of the three representations as representation vectors of graph change information, and determine the second representation of the three representations as representation vectors of graph invariant information.

[0116] In conjunction with the above embodiments, in one implementation, the present invention also provides a method for processing out-of-distribution generalized graphs based on decoupled representation learning. In this method, the graph neural network model to be trained includes at least: a normalizer; the step S15 above, "under the various distribution environments, using the representation vectors of the graph invariant information to perform self-supervised comparative learning to obtain a trained graph neural network model," specifically includes steps S31 to S33:

[0117] Step S31: Through the normalizer, the representation vector of the graph invariant information is made to maintain the stability of the graph representation prediction in the various distribution environments.

[0118] In this embodiment, the normalizer in the graph neural network model to be trained can be used to ensure that the graph invariant information representation vector maintains stability in graph representation prediction under simulated multiple distribution environments. Specifically, the graph invariant information representation vector in this embodiment represents the information in unlabeled graph structure data that truly has graph representation prediction value.

[0119] Step S32: Perform graph representation prediction based on the representation vector of the graph invariant information, and calculate the loss value corresponding to the graph neural network model to be trained through self-supervised contrastive learning.

[0120] In this embodiment, the graph neural network model can focus more on capturing information in graph structure data that truly has graph representation prediction value, while excluding irrelevant graph change information. Graph representation prediction is performed based on the representation vector of graph invariant information to obtain the graph representation prediction result. Through self-supervised contrastive learning, the loss value corresponding to the graph neural network model to be trained is calculated based on the graph representation prediction result.

[0121] Step S33: Based on the hypothesis loss value corresponding to the graph information classification neural network and the loss value corresponding to the graph neural network model to be trained, adjust the model parameters of the graph neural network model to be trained to obtain the trained graph neural network model.

[0122] In this embodiment, during the training process of the model, the graph neural network model to be trained and the graph information classification neural network are trained together. In each round of training, the model parameters of the graph neural network model to be trained and the network parameters of the graph information classification neural network are adjusted based on the hypothesis loss value corresponding to the graph information classification neural network and the loss value corresponding to the graph neural network model to be trained, until the trained graph neural network model and the trained graph information classification neural network are obtained.

[0123] This embodiment can help improve the efficiency of networks (such as graph neural network models and graph information classification neural networks) in recognizing and utilizing core data features.

[0124] In one embodiment, following the above examples, after obtaining a trained graph neural network model and a trained graph information classification neural network, during graph processing, graph structure data (such as a molecular graph) can be input into the trained graph neural network model to obtain an updated node representation for each node in the graph structure data. Then, the updated node representations are processed through K independent graph convolution channels of the trained graph neural network model to obtain K representations for each node in the graph structure data. Next, the trained graph information classification neural network processes the K representations of each node in the graph structure data into graph-invariant information representation vectors and graph-change information representation vectors. Then, based on the graph change information representation vectors determined by the trained graph information classification neural network, the trained graph neural network model simulates various distribution environments. Finally, the trained graph neural network model uses the graph-invariant information representation vectors to predict graph representations under these various distribution environments, obtaining the graph representation of the graph structure data output by the trained graph neural network model. This graph representation of the graph structure data is a deconstructed graph structure data representation, thereby enabling tasks such as classification and regression based on the graph representation of the graph structure data.

[0125] In conjunction with the above embodiments, the present invention also provides a method for processing out-of-distribution generalization graphs based on decoupled representation learning. In this method, the step S14 above, "simulating multiple distribution environments of the unlabeled graph structure data according to the representation vector of the graph change information," specifically includes steps S41 and S42:

[0126] Step S41: Perform clustering based on the representation vector of the graph change information to obtain the clustering result.

[0127] In this embodiment, after obtaining the graph change information, clustering can be performed based on the representation vector of the graph change information to obtain the clustering result.

[0128] Step S42: Based on the clustering results, simulate the various distribution environments to encourage the representation vector of the graph change information to maintain predictability of the distribution environment in the various distribution environments.

[0129] In this embodiment, clustering results can be obtained by clustering the representation vectors of graph change information, simulating various distribution environments of unlabeled graph structure data, so as to encourage the representation vectors of graph change information to maintain predictability of the distribution environment in diverse distribution environments.

[0130] like Figure 2 As shown, Figure 2 This is a schematic flowchart illustrating a distributed external generalization graph pre-training method based on decoupled representation, as shown in an embodiment of the present invention. Figure 2 First, the unlabeled graph structure data is decoupled and encoded using a graph neural network model to be trained, which encodes graph-invariant and graph-change information: For each input graph structure data, the node information of each input graph structure data is aggregated with neighbors, and the information of the current node is updated using the neighbor information of each node to obtain the updated node representation. Through L iterations, all information of the L-order neighbors of each node is aggregated. Then, through K independent graph convolution channels, all information of the L-order neighbors of each node is aggregated independently, thereby decoupling the hidden factors behind the graph structure data.

[0131] Secondly, self-supervised contrastive clustering based on graph change information is employed: A graph information classification neural network, drawing on group theory, identifies graph-invariant and graph-change information and converts them into corresponding representation vectors. Clustering is then performed based on these vector representations to simulate various distribution environments, encouraging graph change information to maintain predictability across these diverse environments. Specifically, clustering can identify multiple environments, allowing for the determination of which information is graph-invariant and which is graph-change information within each environment.

[0132] Then, based on graph invariant information, stable learning in multiple environments is achieved: a regularizer is introduced into the graph neural network model to be trained to enable graph invariant information to maintain the stability of prediction in multiple distribution environments. This allows the graph neural network model to focus more on capturing the information with real predictive value in the graph structure when making graph representation predictions, while excluding irrelevant graph change information, which helps to improve the network's efficiency in recognizing and utilizing core data features.

[0133] Finally, batch training optimization is performed: In each training process, a batch of graph-structured data is selected, input into the graph neural network model to be trained, and prediction results are obtained. Then, through self-supervised contrastive learning, the loss function is calculated, the gradient of the graph neural network model parameters is calculated based on the obtained loss value, and gradient updates are performed to optimize network performance, resulting in a trained graph neural network model. This trained graph neural network model can adaptively generate stable graph representations on out-of-distribution graph-structured data, without relying on label information.

[0134] Ultimately, the self-supervised pre-trained graph neural network model can output pre-trained vectorized representations that generalize out of distribution. These representations can be applied to the testing phase of downstream tasks, such as processing graph datasets in downstream tasks, obtaining representations of each graph, and performing tasks such as classification and regression based on these representations.

[0135] In conjunction with the above embodiments, the present invention also provides a method for processing out-of-distribution generalization graphs based on decoupled representation learning. In this method, in addition to the steps described above, step S51 may also be included, and step S16 may specifically include steps S52 and S53:

[0136] Step S51: Use the labeled molecular graph to test the trained graph neural network model and obtain the test results.

[0137] In this embodiment, the graph dataset includes at least a drug dataset, and the molecular graph includes at least a drug molecular graph. Since insufficient labeled data is often a challenge in the early stages of drug development, after obtaining a trained graph neural network model, a small number (e.g., within a first value, where the first value can be freely set according to requirements) of labeled molecular graphs can be used to test the trained graph neural network model and obtain test results.

[0138] Step S52: If the test results meet the conditions, input the unlabeled drug molecule graphs in the drug dataset into the trained graph neural network model to obtain the graph representation of the drug molecule graphs.

[0139] In this embodiment, if the test results meet certain conditions, it is determined that the trained graph neural network model can be applied to the classification of larger-scale unlabeled drug datasets. At this point, the unlabeled drug molecule graphs from the drug dataset can be input into the trained graph neural network model to obtain graph representations of the drug molecule graphs. The conditions in this embodiment refer to the test results meeting the model's requirements, such as a prediction accuracy greater than a certain value, and are not subject to any restrictions.

[0140] Step S53: Classify the drug dataset based on the graph representation of the drug molecule graph.

[0141] In this embodiment, drug molecule graphs in a drug dataset can be classified based on the graph representation of drug molecule graphs output by a trained graph neural network model. This significantly improves the efficiency and accuracy of drug molecule screening by enhancing the processing capability of unlabeled data, accelerating the discovery of new drugs and reducing R&D costs.

[0142] In conjunction with the above embodiments, the present invention also provides a method for processing out-of-distribution generalization graphs based on decoupled representation learning. In this method, in addition to the steps described above, steps S61 and S62 may also be included:

[0143] Step S61: Input the social network graph data into the trained graph neural network model to obtain the graph representation of the social network graph data.

[0144] In this embodiment, the trained graph neural network model can also be used for social network analysis: social network graph data can be input into the trained graph neural network model to obtain a graph representation of the social network graph data.

[0145] Step S62: Based on the graph representation of the social network graph data, analyze the social network graph data to obtain the analysis results.

[0146] In this embodiment, the social network graph data can be analyzed based on the graph representation of the social network graph data to obtain analysis results. The analysis results include at least the dynamic changes and data migration of the social network corresponding to the social network graph data.

[0147] Considering that social network data is typically dynamic and has distributed migration characteristics, traditional models struggle to adapt to environmental changes. Therefore, this embodiment uses a pre-trained graph neural network model to predict an environment-adaptive graph representation, which can effectively handle dynamic changes and data migration, provide stable and generalized analysis results, enhance the accuracy of user behavior prediction, and improve user experience.

[0148] In conjunction with the above embodiments, the present invention also provides a method for processing out-of-distribution generalization graphs based on decoupled representation learning. In this method, in addition to the steps described above, steps S71 and S72 may also be included:

[0149] Step S71: Input the graph data in the human-computer interaction system into the trained graph neural network model to obtain the graph representation corresponding to the human-computer interaction.

[0150] Considering that accurately understanding and predicting user behavior is crucial in human-computer interaction systems, this embodiment can also use the trained graph neural network model for human-computer interaction and computer-aided systems. The graph data in the human-computer interaction system is input into the trained graph neural network model to obtain the graph representation of the human-computer interaction output.

[0151] Step S72: Based on the graph representation corresponding to the human-computer interaction, understand and predict the user needs and user behaviors in the human-computer interaction system.

[0152] In this embodiment, user needs and behaviors in the human-computer interaction system can be understood and predicted based on the graph representation corresponding to human-computer interaction. The stability and generalization ability of the graph neural network model trained in this embodiment enable the human-computer interaction system to understand user needs and behavior patterns more accurately, thereby providing a more personalized and efficient interactive experience.

[0153] In summary, the out-of-distribution generalization graph processing method based on decoupled representation learning proposed in this invention enables graph neural network models to achieve relatively stable performance even with shifted distributions of training and testing data through self-supervised learning, without graph labels. In other words, even when training and testing environments are inconsistent, the method in this embodiment can adaptively and self-supervisedly compute and generate generalized graph structure data representations for out-of-distribution data without relying on environmental labels, thereby achieving more reliable and accurate application results.

[0154] It should be noted that, for the sake of simplicity, the method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0155] Based on the same inventive concept, one embodiment of the present invention provides a distributed out-of-generalization graph processing system based on decoupled representation learning. (Reference) Figure 3 , Figure 3 This is a structural block diagram of a distributed external generalization graph processing system based on decoupled representation learning, as shown in an embodiment of the present invention. Figure 3 As shown, the system 300 includes:

[0156] The node representation module 301 is used to input unlabeled graph structure data into the graph neural network model to be trained, and adopt a neighbor aggregation mechanism to update the node representation of each node in the unlabeled graph structure data through L iterations based on the neighbor information to obtain the updated node representation; wherein, the updated node representation of a node aggregates all the information of the node's L-order neighbors.

[0157] The representation processing module 302 is used to process the updated node representations through the K independent graph convolution channels of the graph neural network model to be trained, so as to obtain K representations for each node; where K and L are natural numbers greater than 1.

[0158] The representation classification module 303 is used to process the K representations of each node into a representation vector of graph invariant information and a representation vector of graph change information;

[0159] The environment simulation module 304 is used to simulate various distribution environments of the unlabeled graph structure data based on the representation vector of the graph change information through the graph neural network model to be trained.

[0160] The model training module 305 is used to perform self-supervised comparative learning using the graph invariant information representation vectors under the various distribution environments through the graph neural network model to be trained, so as to obtain a trained graph neural network model.

[0161] Graph processing module 306 is used to input molecular graphs in the graph dataset into the trained graph neural network model to obtain graph representations of the molecular graphs, and to classify the graph dataset based on the graph representations of the molecular graphs.

[0162] Optionally, the characterization and classification module 303 includes:

[0163] The first assumption module is used to classify neural networks through graph information. Based on the idea of ​​group theory, it assumes that any one or more of the K representations are representation vectors of candidate graph change information, and assumes that the remaining representations of the K representations are representation vectors of candidate graph invariant information.

[0164] The environment hypothesis module is used to simulate multiple hypothesis distribution environments based on the representation vector of the candidate graph change information;

[0165] The first loss calculation module is used to perform self-supervised comparative learning using the representation vector of the candidate graph invariant information under the multiple hypothesis distribution environments, and calculate the hypothesis loss value corresponding to the graph information classification neural network.

[0166] The second hypothesis module is used to re-hypothesize any one or more of the K representations as the representation vector of the candidate graph change information, and to assume that the remaining representations of the K representations are the representation vector of the candidate graph invariant information, until the hypothesis loss value of all hypothetical cases of the K representations is calculated.

[0167] The representation determination module is used to process the K representations corresponding to the minimum hypothesis loss value into a representation vector of the graph invariant information and a representation vector of the graph change information, based on the hypothesis of the K representations.

[0168] Optionally, the environment simulation module 304 includes:

[0169] The clustering module is used to perform clustering based on the representation vector of the graph change information to obtain clustering results;

[0170] An environment determination module is used to simulate the various distribution environments based on the clustering results, so as to encourage the representation vector of the graph change information to maintain predictability of the distribution environment in the various distribution environments.

[0171] Optionally, the graph neural network model to be trained includes at least: a regularizer; and a model training module 305, including:

[0172] The normalization module is used to ensure that the representation vector of the graph invariant information maintains stability in graph representation prediction in various distribution environments through the normalizer. The representation vector of the graph invariant information represents the information in the unlabeled graph structure data that truly has graph representation prediction value.

[0173] The second loss calculation module is used to perform graph representation prediction based on the representation vector of the graph invariant information, and calculate the loss value corresponding to the graph neural network model to be trained through self-supervised contrastive learning.

[0174] The model parameter tuning module is used to adjust the model parameters of the graph neural network model to be trained based on the hypothesis loss value corresponding to the graph information classification neural network and the loss value corresponding to the graph neural network model to be trained, so as to obtain the trained graph neural network model.

[0175] Optionally, the graph dataset includes a drug dataset, the molecular graph includes a drug molecular graph, and the system 300 further includes:

[0176] The model testing module is used to test the trained graph neural network model using labeled molecular graphs and obtain test results.

[0177] The graph processing module 306 includes:

[0178] The drug characterization module is used to input the unlabeled drug molecule graphs in the drug dataset into the trained graph neural network model to obtain the graph representation of the drug molecule graphs when the test results meet the conditions.

[0179] The drug classification module is used to classify the drug dataset based on the graph representation of the drug molecule graph.

[0180] Optionally, the system 300 further includes:

[0181] The social data processing module is used to input social network graph data into the trained graph neural network model to obtain a graph representation of the social network graph data.

[0182] The social data analysis module is used to analyze the social network graph data based on the graph representation of the social network graph data and obtain analysis results.

[0183] The analysis results include at least the dynamic changes and data migration of the social networks corresponding to the social network graph data.

[0184] Optionally, the system 300 further includes:

[0185] The human-computer interaction processing module is used to input the graph data of the human-computer interaction system into the trained graph neural network model to obtain the graph representation corresponding to the human-computer interaction.

[0186] The human-computer interaction analysis module is used to understand and predict user needs and user behaviors in the human-computer interaction system based on the graph representation corresponding to the human-computer interaction.

[0187] Based on the same inventive concept, another embodiment of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in the distribution out-generalization graph processing method based on decoupled representation learning as described in any of the above embodiments of the present invention.

[0188] Based on the same inventive concept, another embodiment of the present invention provides an electronic device 400, such as... Figure 4 As shown. Figure 4 This is a schematic diagram of an electronic device according to an embodiment of the present invention. The electronic device includes a memory 402, a processor 401, and a computer program stored in the memory and executable on the processor. When executed by the processor, the program implements the steps in the distributed out-generalization graph processing method based on decoupled representation learning described in any of the above embodiments of the present invention.

[0189] As the system implementation is basically similar to the method implementation, it is described in a relatively simple way. For relevant details, please refer to the description of the system implementation.

[0190] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0191] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0192] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0193] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0194] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0195] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present invention.

[0196] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0197] The above provides a detailed description of the distribution out-generalization graph processing method, system, and product based on decoupled representation learning provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for processing out-of-distribution generalization graphs based on decoupled representation learning, characterized in that, The method includes: Unlabeled graph structure data is input into the graph neural network model to be trained. The neighbor aggregation mechanism is used to update the node representation of each node in the unlabeled graph structure data L times according to the neighbor information to obtain the updated node representation. The updated node representation of a node aggregates all the information of the node's L-order neighbors. The updated node representations are processed through the K independent graph convolution channels of the graph neural network model to be trained to obtain K representations for each node; where K and L are natural numbers greater than 1. The K representations of each node are processed into a representation vector of graph-invariant information and a representation vector of graph-change information; The graph neural network model to be trained simulates various distribution environments of the unlabeled graph structure data based on the representation vector of the graph change information. Using the graph neural network model to be trained, in the various distribution environments, self-supervised comparative learning is performed using the representation vector of the graph invariant information to obtain a trained graph neural network model. The molecular graphs in the graph dataset are input into the trained graph neural network model to obtain the graph representations of the molecular graphs. The graph dataset is then classified based on the graph representations of the molecular graphs. The graph dataset includes a drug dataset, and the molecular graphs include drug molecular graphs. The process of processing the K representations of each node into a representation vector of graph-invariant information and a representation vector of graph-transformative information includes: Based on the idea of ​​group theory, the graph information classification neural network assumes that any one or more of the K representations are the representation vectors of candidate graph change information, and assumes that the remaining representations of the K representations are the representation vectors of candidate graph invariant information. Based on the representation vector of the candidate graph change information, various hypothetical distribution environments are simulated; Under the various hypothetical distribution environments, self-supervised comparative learning is performed using the representation vector of the candidate graph invariant information to calculate the hypothesis loss value corresponding to the graph information classification neural network. Re-assume that any one or more of the K representations are representation vectors of the candidate graph change information, and assume that the remaining representations of the K representations are representation vectors of the candidate graph invariant information, until the assumption loss values ​​of all assumption cases of the K representations are calculated; Based on the assumptions of the K representations corresponding to the minimum hypothesis loss value, the K representations are processed into representation vectors of graph invariant information and representation vectors of graph change information.

2. The method according to claim 1, characterized in that, The representation vector based on the graph change information is used to simulate various distribution environments of the labelless graph structure data, including: Clustering is performed based on the representation vectors of the graph change information to obtain the clustering results; Based on the clustering results, the various distribution environments are simulated to encourage the representation vectors of graph change information to maintain predictability of the distribution environment in the various distribution environments.

3. The method according to claim 1, characterized in that, The graph neural network model to be trained includes at least: a normalizer; under the various distribution environments, self-supervised contrastive learning is performed using the representation vectors of the graph-invariant information to obtain a trained graph neural network model, including: The normalizer enables the graph-invariant information representation vector to maintain stability in graph representation prediction across various distribution environments. The graph-invariant information representation vector represents the information in the unlabeled graph structure data that truly has graph representation prediction value. Graph representation prediction is performed based on the representation vector of the graph invariant information, and the loss value corresponding to the graph neural network model to be trained is calculated through self-supervised contrastive learning. Based on the hypothetical loss value corresponding to the graph information classification neural network and the loss value corresponding to the graph neural network model to be trained, the model parameters of the graph neural network model to be trained are adjusted to obtain the trained graph neural network model.

4. The method according to any one of claims 1 to 3, characterized in that, The method further includes: The trained graph neural network model was tested using labeled molecular graphs to obtain test results; The step of inputting molecular graphs from the graph dataset into the trained graph neural network model to obtain graph representations of the molecular graphs, and classifying the graph dataset based on the graph representations of the molecular graphs, includes: If the test results meet the conditions, the unlabeled drug molecule graphs in the drug dataset are input into the trained graph neural network model to obtain the graph representation of the drug molecule graphs. The drug dataset is classified based on the graph representation of the drug molecule graph.

5. The method according to any one of claims 1 to 3, characterized in that, The method further includes: The social network graph data is input into the trained graph neural network model to obtain the graph representation of the social network graph data. Based on the graph representation of the social network graph data, the social network graph data is analyzed to obtain analysis results; The analysis results include at least the dynamic changes and data migration of the social networks corresponding to the social network graph data.

6. The method according to any one of claims 1 to 3, characterized in that, The method further includes: The graph data in the human-computer interaction system is input into the trained graph neural network model to obtain the graph representation corresponding to the human-computer interaction. Based on the graph representation corresponding to the human-computer interaction, user needs and user behaviors in the human-computer interaction system are understood and predicted.

7. A distributed out-generalization graph processing system based on decoupled representation learning, characterized in that, The system includes: The node representation module is used to input unlabeled graph structure data into the graph neural network model to be trained. It adopts a neighbor aggregation mechanism to update the node representation of each node in the unlabeled graph structure data through L iterations based on the neighbor information to obtain the updated node representation. The updated node representation of a node aggregates all the information of the node's L-order neighbors. The representation processing module is used to process the updated node representations through K independent graph convolution channels of the graph neural network model to be trained, so as to obtain K representations for each node; where K and L are natural numbers greater than 1. The representation classification module is used to process the K representations of each node into a representation vector of graph-invariant information and a representation vector of graph-change information; The environment simulation module is used to simulate various distribution environments of the unlabeled graph structure data based on the representation vector of the graph change information through the graph neural network model to be trained. The model training module is used to perform self-supervised comparative learning using the graph invariant information representation vectors of the graph neural network model to be trained under the various distribution environments, so as to obtain a trained graph neural network model. The graph processing module is used to input molecular graphs from the graph dataset into the trained graph neural network model to obtain graph representations of the molecular graphs, and to classify the graph dataset based on the graph representations of the molecular graphs; the graph dataset includes a drug dataset, and the molecular graphs include drug molecule graphs; The characterization and classification module includes: The first assumption module is used to classify neural networks through graph information. Based on the idea of ​​group theory, it assumes that any one or more of the K representations are representation vectors of candidate graph change information, and assumes that the remaining representations of the K representations are representation vectors of candidate graph invariant information. The environment hypothesis module is used to simulate multiple hypothesis distribution environments based on the representation vector of the candidate graph change information; The first loss calculation module is used to perform self-supervised comparative learning using the representation vector of the candidate graph invariant information under the multiple hypothesis distribution environments, and calculate the hypothesis loss value corresponding to the graph information classification neural network. The second hypothesis module is used to re-hypothesize any one or more of the K representations as the representation vector of the candidate graph change information, and to assume that the remaining representations of the K representations are the representation vector of the candidate graph invariant information, until the hypothesis loss value of all hypothetical cases of the K representations is calculated. The representation determination module is used to process the K representations corresponding to the minimum hypothesis loss value into a representation vector of the graph invariant information and a representation vector of the graph change information, based on the hypothesis of the K representations.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the computer program is executed by the processor, it implements the distributional out-generalization graph processing method based on decoupled representation learning as described in any one of claims 1-6.

9. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by the processor, it implements the distributional out-generalization graph processing method based on decoupled representation learning as described in any one of claims 1-6.

Images