Model training method, device, equipment and storage medium

Abstract

The present application provides a model training method, apparatus, device and storage medium, the method comprising: obtaining first training graph data and N groups of hyperparameters, the first training graph data comprising R data pairs, each data pair consisting of the first feature information of a central node in the graph network and the first feature information of a neighboring graph node of the central node; inputting the first training graph data into the generator under each set of hyperparameters in the N groups of hyperparameters, respectively, so that the generator learns the feature probability distribution of the neighboring nodes of the central node under the given feature conditions of the central node, and obtains the generator trained under each set of hyperparameters in the N groups of hyperparameters; determining a target generator from the generators trained under the N groups of hyperparameters, the target generator being used to generate a second sampling sample, and the second sampling sample being used to train a preset graph neural network model, thereby solving the problem of overfitting of the model due to lack of samples during training, and improving the robustness and generalization of the model.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of artificial intelligence technology, and in particular to a model training method, apparatus, device, and storage medium. Background Art

[0002] With the rapid development of artificial intelligence (AI), various graph neural network (GNN) models have emerged. The prediction accuracy of GNN models is closely related to the training process, which in turn is closely related to sample preparation. For example, insufficient sample space or sample quantity can severely impact training or lead to insufficient generalization of the trained model, resulting in low recognition and accuracy rates.

[0003] Therefore, how to solve the problem of insufficient generalization of model training caused by insufficient number of samples has become a technical problem that needs to be urgently solved by those skilled in the art. Summary of the Invention

[0004] The present application provides a model training method, apparatus, device and storage medium. The trained generator can generate sampling samples for model training to achieve sufficient training of the model.

[0005] In a first aspect, the present application provides a model training method, comprising:

[0006] Obtain first training graph data and N sets of hyperparameters, where the first training graph data includes R data pairs, each data pair consisting of first feature information of a central node in the graph network and first feature information of a neighboring graph node of the central node, where N and R are both positive integers;

[0007] Inputting the first training graph data into the generator under each set of hyperparameters in the N sets of hyperparameters respectively, so that the generator learns the feature probability distribution of the neighboring nodes of the central node under the given feature conditions of the central node, and obtaining the generator trained under each set of hyperparameters in the N sets of hyperparameters;

[0008] A target generator is determined from the generators trained under N sets of hyperparameters, and the target generator is used to generate a second sampling sample, and the second sampling sample is used to train a preset graph neural network model.

[0009] In a second aspect, a model training device is provided, comprising:

[0010] an acquisition unit, configured to acquire first training graph data and N sets of hyperparameters, wherein the first training graph data includes R data pairs, each data pair consisting of first feature information of a central node in a graph network and first feature information of a neighboring graph node of the central node, and N and R are both positive integers;

[0011] A training unit is configured to input the first training graph data into a generator under each of the N sets of hyperparameters, so that the generator learns the feature probability distribution of the neighboring nodes of the central node under the given feature conditions of the central node, and obtains a generator trained under each of the N sets of hyperparameters;

[0012] A determination unit is used to determine a target generator from the generators trained under N groups of hyperparameters, wherein the target generator is used to generate a second sampling sample, and the second sampling sample is used to train a preset graph neural network model.

[0013] In some embodiments, the determination unit is specifically used to sample each set of hyperparameters in the N sets of hyperparameters using the generator trained under the set of hyperparameters to obtain a first sampling sample output by the generator trained under the set of hyperparameters, and use the first sampling sample output by the generator trained under the set of hyperparameters to train a preset first prediction model to obtain the trained first prediction model corresponding to the set of hyperparameters; determine the prediction accuracy of the trained first prediction model corresponding to each set of hyperparameters respectively; and determine the generator trained under a set of hyperparameters corresponding to the first prediction model with the highest prediction accuracy as the target generator.

[0014] In some embodiments, the training unit is specifically used to obtain an i-th batch from the first training graph data for the generator under each group of hyperparameters in the N groups of hyperparameters according to a preset batch size, wherein the i-th batch includes at least one data pair, wherein i is a positive integer from 1 to M, and M is a preset number of training times of the generator; the generator under the group of hyperparameters is trained using the i-th batch so that the generator learns the feature probability distribution of the neighboring nodes of each central node under the given feature conditions of each central node in the i-th batch, and obtains the generator after the i-th batch training; and the generator is trained using The generator after training the i-th batch is sampled to obtain the third sampling sample corresponding to the i-th batch output by the generator; the third sampling sample corresponding to the i-th batch is input into the preset second prediction model to obtain the prediction result corresponding to the i-th batch output by the second prediction model; according to the prediction result corresponding to the i-th batch output by the second prediction model, the uncertainty score corresponding to the i-th batch is determined; the generator trained by the batch corresponding to the highest uncertainty score among the uncertainty scores corresponding to M batches is determined as the generator trained under this group of hyperparameters.

[0015] In some embodiments, the generator is a conditional variational autoencoder, which includes an encoding module and a decoding module, and a training unit, which is specifically used to input the data pair into the encoding module for each data pair in the i-th batch to obtain the first variance and the first mean output by the encoding module; convert the first variance and the first mean into a first latent variable through a reparameterization technique, and input the first latent variable and the first feature information of the central node in the data pair into the decoding module to obtain the augmented feature vector output by the decoding module; use the augmented feature vector to train the encoding module and the decoding module to obtain the encoding module and the decoding module after training the i-th batch.

[0016] In some embodiments, the training unit is specifically used to select the first feature information of K nodes, where K is a positive integer; randomly select a second variance and a second mean for each of the K nodes; for each of the K nodes, convert the second variance and second mean corresponding to the node into a second latent variable through a reparameterization technique, and input the second latent variable and the first feature information of the node into the decoding module after the i-th batch training to obtain a third sampling sample output by the decoding module.

[0017] In some embodiments, the training unit is further used to determine whether the uncertainty score corresponding to the i-th batch is the highest uncertainty score among the uncertainty scores corresponding to the previous i batches; if it is determined that the uncertainty score corresponding to the i-th batch is the highest uncertainty score among the uncertainty scores corresponding to the previous i batches, the second prediction model is trained a first number of times using the third sampling sample output by the generator after training the i-th batch; wherein the second prediction model after training the first number of times is used to predict the prediction result of the third sampling sample output by the generator after training the i+1-th batch.

[0018] In some embodiments, if i=1, in some embodiments, the training unit is also used to pre-train the second prediction model a second number of times; the third sampling sample corresponding to the first batch is input into the second prediction model after the second number of training times to obtain the prediction result corresponding to the first batch output by the second prediction model.

[0019] In some embodiments, the training unit is further used to use the target generator to perform sampling to obtain a second sampling sample output by the target generator; and use the second sampling sample output by the target generator to train a preset graph neural network model.

[0020] In some embodiments, the generator is a conditional variational autoencoder, which includes a decoding module and a training unit, specifically used to obtain second training graph data, wherein the second training graph data includes first feature information of multiple second nodes; randomly selects a third difference and a third mean for each of the multiple second nodes; for each second node in the multiple second nodes, converts the third difference and the third mean corresponding to the second node into a third latent variable through a reparameterization technique, and inputs the third latent variable and the first feature information of the second node into the target decoding module in the target generator to obtain P second sampling samples about the second node output by the target decoding module, where P is a positive integer.

[0021] In some embodiments, the training unit is specifically used to input the second training graph data and the second sampling sample output by the target decoding module into the graph neural network model to train the graph neural network model.

[0022] In some embodiments, the graph neural network model includes a first feature extraction unit, P second feature extraction units and a prediction unit, and a training unit, which is specifically used to input the first feature information of each second node among the multiple second nodes into the first feature extraction unit to obtain the second feature information of the first node output by the first feature extraction unit; input the P second sampling samples of the second node into the P second feature extraction units respectively to obtain the second feature information of each second sampling sample in the P second sampling samples; cascade the second feature information of the second node and the second feature information of the P second sampling samples to obtain the cascaded second feature information; input the cascaded second feature information into the prediction unit to train the graph neural network model.

[0023] In a third aspect, a computing device is provided, comprising a processor and a memory, wherein the memory is configured to store a computer program, and the processor is configured to call and execute the computer program stored in the memory to perform the method of the first aspect or its implementations.

[0024] In a fourth aspect, a chip is provided for implementing the method described in any one of the first and second aspects above, or in their respective implementations. Specifically, the chip includes a processor configured to load and execute a computer program from a memory, causing a device equipped with the chip to perform the method described in any one of the first aspect above, or in their respective implementations.

[0025] In a fifth aspect, a computer-readable storage medium is provided for storing a computer program, wherein the computer program enables a computer to execute the method of any aspect of the first aspect or its various implementations.

[0026] In a sixth aspect, a computer program product is provided, comprising computer program instructions, wherein the computer program instructions enable a computer to execute the method of any one of the first aspects or its various implementations.

[0027] In a seventh aspect, a computer program is provided, which, when executed on a computer, enables the computer to execute the method in any one of the above-mentioned first aspects or its various implementations.

[0028] In summary, the present application obtains the first training graph data and N groups of hyperparameters, wherein the first training graph data includes R data pairs, each data pair consists of the first feature information of a central node in the graph network and the first feature information of a neighboring graph node of the central node, and N and R are both positive integers; the first training graph data is input into the generator under each set of hyperparameters in the N groups of hyperparameters, so that the generator learns the feature probability distribution of the neighboring nodes of the central node under the given feature conditions of the central node, and obtains the generator trained under each set of hyperparameters in the N groups of hyperparameters; from the generator trained under the N groups of hyperparameters, a target generator is determined, and the target generator is used to generate a second sampling sample, and the second sampling sample is used to train a preset graph neural network model. That is, the generator of the embodiment of the present application uses the first feature of the central node on the graph as a condition to learn the feature probability distribution of the neighboring nodes near the central node, so that the generator not only learns the topological structure information of the graph, but also learns the feature information of the graph nodes. Using this generator for data augmentation solves the problem of overfitting of the model due to lack of sufficient samples during training, greatly improving the robustness and generalization of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0030] Figure 1 A schematic diagram of a system architecture involved in an embodiment of the present application;

[0031] Figure 2 A flowchart of a model training method provided in one embodiment of the present application;

[0032] Figure 3 A schematic diagram of a model training process involved in an embodiment of the present application;

[0033] Figure 4 Schematic diagram of a conditional variational autoencoder according to an embodiment of the present application;

[0034] Figure 5 A schematic diagram of a training process involved in an embodiment of the present application;

[0035] Figure 6 A schematic diagram of a network model involved in an embodiment of the present application;

[0036] Figure 7 A flowchart of a model training method provided in one embodiment of the present application;

[0037] Figure 8 A schematic diagram of model training provided in one embodiment of the present application;

[0038] Figure 9 Schematic diagram of the model training process provided in one embodiment of the present application

[0039] Figure 10 is a schematic block diagram of a model training device provided in one embodiment of the present application;

[0040] Figure 11 It is a schematic block diagram of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0042] It should be understood that in the embodiments of the present invention, "B corresponding to A" means that B is associated with A. In one implementation, B can be determined based on A. However, it should also be understood that determining B based on A does not mean determining B based solely on A; B can also be determined based on A and / or other information.

[0043] In the description of the present application, unless otherwise specified, "plurality" means two or more than two.

[0044] In addition, to facilitate the clear description of the technical solutions of the embodiments of the present application, in the embodiments of the present application, the words "first" and "second" are used to distinguish between identical or similar items with substantially the same functions and effects. Those skilled in the art will understand that the words "first" and "second" do not limit the quantity or execution order, and the words "first" and "second" do not necessarily mean different.

[0045] To facilitate understanding of the embodiments of the present application, the following briefly introduces the relevant concepts involved in the embodiments of the present application:

[0046] Artificial Intelligence (AI) refers to the theories, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0047] Artificial intelligence (AI) technology is a comprehensive discipline encompassing a wide range of fields, encompassing both hardware and software technologies. Foundational AI technologies generally include sensors, specialized AI chips, cloud computing, distributed storage, big data processing, operating / interaction systems, and mechatronics. AI software technologies primarily encompass computer vision, speech processing, natural language processing, and machine learning / deep learning.

[0048] Machine learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and self-learning.

[0049] Natural language processing (NLP) is a key area of ​​research in computer science and artificial intelligence. It studies the theories and methods that enable effective communication between humans and computers using natural language. Natural language processing (NLP) integrates linguistics, computer science, and mathematics. Therefore, research in this field involves natural language—the language we use in everyday life—and is closely linked to the study of linguistics. Natural language processing technologies typically include text processing, semantic understanding, machine translation, robotic question answering, and knowledge graphs.

[0050] Natural language models: These models use statistical models to convert large amounts of human language into machine language for cognition, understanding, and generation. Specific applications include machine translation and automated question-answering.

[0051] Pre-training: A language model is trained on a large amount of unlabeled text to obtain a set of model parameters. This set of parameters is used to initialize the model, achieving a "warm start." Based on the specific task, the parameters are then fine-tuned within the existing language model architecture to fit the labeled data provided by the task. Pre-training has proven effective in both classification and tagging tasks in natural language processing.

[0052] Graph neural networks (Graph Neural Networks): These fall under the umbrella of deep learning and are commonly used to solve a range of graph-based problems, such as graph classification, graph node classification, and graph edge prediction. Graph neural networks consist of two parts: an adjacency matrix that describes the graph structure and a feature matrix that describes the graph attributes. The second part embeds the graph structure and attributes into the learning process, and can be a feedforward neural network or an attention-based neural network.

[0053] Conditional variational autoencoder (CVA): A conditional variational autoencoder is a generative network that uses an encoder to learn the data distribution, generating latent variables. A decoder then restores these latent variables to their original form. This network can learn the data distribution and then sample new data. It is commonly used for image and text generation.

[0054] Graph convolutional neural networks (GCNNs): A type of GNN, GCNNs operate on the features of nearby nodes using a neighborhood matrix. They weight the feature vectors of all nodes locally to obtain the feature vector of the central node. GCNNs can perform smooth operations on the features of all nodes on a graph, facilitating downstream node classification tasks.

[0055] Importance sampling: Importance sampling is a selective sampling of data. It uses a neural network to score the data based on its uncertainty and then extracts the data with the highest scores. This method can increase the diversity of the neural network's input data, exposing the neural network to more diverse data and improving generalization.

[0056] Concatenation: Deep learning performs cascade operations on latent variables between networks, which can maximize the distribution of different variables and will not destroy the distribution of data by averaging the latent variables.

[0057] In some embodiments, the system architecture of the embodiment of the present application is as follows Figure 1 shown.

[0058] Figure 1 This is a schematic diagram of a system architecture involved in an embodiment of the present application, including a user device 101, a data acquisition device 102, a training device 103, an execution device 104, a database 105 and a content library 106.

[0059] The data acquisition device 102 is configured to read training data from the content library 106 and store the read training data in the database 105. The training data involved in the embodiment of the present application includes first training graph data including R data pairs, where one data pair consists of first feature information of a central node in the graph structure and first feature information of a neighboring graph node of the central node.

[0060] In some embodiments, the user device 101 is used to perform a labeling operation on the data in the database 105 , for example, labeling the classification of each node.

[0061] The training device 103 trains the generator based on the training data maintained in the database 105, so that the trained generator can sample a second sample, which can be used to train other preset graph neural network models. Optionally, the generator obtained by the training device 103 can be applied to different systems or devices.

[0062] In the attached Figure 1 In the example, the execution device 104 is configured with an I / O interface 107 for data exchange with external devices. For example, the I / O interface receives the second training graph data sent by the user device 101. The computing module 109 in the execution device 104 processes the input second training graph data using the trained generator, outputs sampled data, and inputs the generated sampled data into the graph neural network model to be trained to train the graph neural network model.

[0063] The user device 101 may include a mobile phone, a tablet computer, a laptop computer, a PDA, a mobile internet device (MID), or other terminal devices with a browser installed.

[0064] The execution device 104 may be a server.

[0065] For example, the server may be a computing device such as a rack server, a blade server, a tower server, or a cabinet server, etc. The server may be an independent test server or a test server cluster composed of multiple test servers.

[0066] In this embodiment, the execution device 104 is connected to the user device 101 via a network. The network can be a wireless or wired network such as an intranet, the Internet, the Global System of Mobile Communication (GSM), Wideband Code Division Multiple Access (WCDMA), a 4G network, a 5G network, Bluetooth, Wi-Fi, or a call network.

[0067] It should be noted that the attached Figure 1 It is only a schematic diagram of a system architecture provided by an embodiment of the present application. The positional relationship between the devices, components, modules, etc. shown in the figure does not constitute any limitation. In some embodiments, the above-mentioned data acquisition device 102 and the user device 101, the training device 103 and the execution device 104 can be the same device. The above-mentioned generator and the graph neural network model can be located on different computing devices. The above-mentioned database 105 can be distributed on one server or on multiple servers, and the above-mentioned content library 106 can be distributed on one server or on multiple servers.

[0068] The embodiment of the present application provides a model training method, in which the first training graph data for training the generator includes R data pairs, and one data pair is composed of the first feature information of a central node in the graph structure and the first feature information of a neighboring graph node of the central node. That is, the generator of the embodiment of the present application uses the first feature of the central node on the graph as a condition to learn the feature probability distribution of the neighboring nodes near the central node, so that the generator not only learns the topological structure information of the graph, but also learns the feature information of the graph nodes. In this way, the trained generator can generate more sampling samples that meet the requirements. When these sampling samples are used to train the graph neural network model, the graph neural network model can be fully trained, and the generalization degree, recognition rate and accuracy of the trained graph neural network model can be improved.

[0069] The application scenarios of this application include but are not limited to any graph deep learning scenarios such as social networks, artificial intelligence pharmaceuticals, recommendation systems, and traffic forecasts. For example, in the recommendation system, there are tasks such as recommending friends and products to users. Since the underlying recommendation mechanisms are the same, the method of the embodiment of this application can be used to extract local substructure information well, better enrich the characteristics of users, improve the quality of recommendations, and learn strategies that can complete multiple tasks at the same time and be generalized to new tasks. For example, in the artificial intelligence pharmaceutical application scenario, the embodiment of this application can use the collected drug data to learn the structure and feature information near different functional groups, summarize the information of the local structure of drug molecules, and improve the efficiency of drug synthesis.

[0070] The following describes the technical solutions of the embodiments of the present application in detail through some embodiments. The following embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0071] Figure 2 This is a flowchart of a model training method provided in one embodiment of the present application. The execution subject of this embodiment of the present application is a device with a model training function, such as a model training device, which can be a computing device or a portion of a computing device. The following description uses the computing device as an example.

[0072] like Figure 2 As shown, the method of the embodiment of the present application includes:

[0073] S201. Obtain first training graph data and N sets of hyperparameters.

[0074] The first training graph data includes R data pairs, each data pair consists of the first feature information of a central node in the graph network and the first feature information of a neighboring graph node of the central node, and N and R are both positive integers.

[0075] Graph data contains a wealth of relational information. Reasoning and learning from unstructured data such as text and images, such as sentence dependency trees and image scene graphs, requires graph reasoning models. Graph neural networks are a connectionist model that captures dependencies within a graph by transferring information between nodes.

[0076] A graph is a structured data structure consisting of a series of objects (nodes) and relationships (edges). As a non-Euclidean data structure, graph analysis has been applied to node classification, link prediction, and clustering.

[0077] The graph can usually be represented as are all nodes {v1,…,v N}, ε is the set of edges, where |V| = N. The adjacency matrix can be defined as A∈{0,1} N×N , and if and only if (v i , v j )∈ε, there is A ij = 1. Let Represented as node v i The neighbor node set of D represents the diagonal matrix, then The feature matrix is ​​expressed as Each node v corresponds to an F-dimensional feature vector X v Y∈{0,1} N×C represents the one-hot label matrix, where Y i ∈{0, 1} C is a one-hot vector and For any v i ∈V holds true.

[0078] The central node described in the embodiment of the present application can be understood as any node v in the figure. i , central node v i The neighbor nodes are the central node v i Nodes belonging to the same edge, that is, the central node v i Connected nodes.

[0079] The first feature information of the node in the embodiment of the present application can be understood as the feature vector X of the node v .

[0080] The first training graph data of the embodiment of the present application includes R data pairs, and one data pair consists of the first feature information of a central node and the first feature information of a neighboring node of the central node.

[0081] For example, taking the central node v2 as an example, the neighbor nodes of the central node v2 include v3 and v4, the first feature information of v2 is recorded as X2, the first feature information of v3 is recorded as X3, and the first feature information of v4 is recorded as X4, among which (X2, X3) constitutes a data pair, and (X2, X4) constitutes a data pair.

[0082] The embodiment of the present application does not limit the number of data pairs R included in the first training graph data, and the number is determined based on actual needs.

[0083] In some embodiments, R is a positive integer greater than or equal to 2.

[0084] During the model training process, model parameters include known parameters and unknown parameters. The model training process can be understood as the process of determining unknown parameters. The known parameters are also called super parameters. In this embodiment, super parameters are referred to as hyper parameters.

[0085] In one example, the N groups of hyperparameters may be preset values.

[0086] In one example, the N groups of hyperparameters are empirical values.

[0087] The embodiment of the present application does not limit the method of obtaining the first training image data.

[0088] In one example, the first training graph data is uploaded by a user.

[0089] In one example, a graph network is input into a computing device, which extracts first feature information of each node in the graph network. Based on the structural relationship and first feature information of each node in the graph network, the computing device combines the first feature information of the central node and the first feature information of a neighboring node of the central node into a data pair, sequentially forming R data pairs, which constitute the first training graph data. Optionally, the nodes corresponding to the above R data pairs are nodes corresponding to a locally connected graph in the graph network. Optionally, the nodes corresponding to the above R data pairs may not be connected.

[0090] S202. Input the first training graph data into the generator under each set of hyperparameters in the N groups of hyperparameters, so that the generator learns the feature probability distribution of the neighboring nodes of the central node under the given feature conditions of the central node, and obtains the generator trained under each set of hyperparameters in the N groups of hyperparameters.

[0091] like Figure 3 As shown, the local augmentation of the embodiment of the present application is to learn the conditional distribution of the surrounding node features given the feature information of the central node. Figure 3 As shown, X v is the feature information (or feature representation or feature vector) of the central node, X u1 、X u2 and X u3 is the feature information (or feature representation or feature vector) of the neighboring nodes of the central node v.

[0092] In the embodiment of the present application, during the learning process of the generator, a single distribution is learned for the neighbors of all nodes using maximum likelihood estimation, that is, solving the optimization problem shown in formula (1):

[0093]

[0094] Among them, the first training graph data is a set where X i The central node vi The first characteristic information, The central node v i Neighbor node v j The first characteristic information. ψ Can be used to augment data for all neighbors. For a local substructure, assume that each neighbor satisfies a different conditional distribution, that is, there exists a conditional distribution with hidden variables z j p(·|X i , z j ),for Make X j ~p(X|X i , z j ), “~” means obey or comply. Once we obtain p(·|X i , z j ), then the augmented features can be generated

[0095] From the above, it can be seen that the first training graph data of the embodiment of the present application includes R data pairs, each data pair includes the first feature information of a central node and the first feature information of a neighbor node of the central node, and the first feature information of the central node and the first feature information of the neighbor node of the central node are respectively input into the generator under each set of hyperparameters, so that the generator learns the feature probability distribution of the neighbor nodes of the central node under the given feature conditions of the central node, that is, the above formula (1) is used as the objective function of the generator, and the generator is trained to obtain the generator trained under each set of hyperparameters in N groups of hyperparameters.

[0096] The embodiment of the present application does not limit the method for training the generator using the first training graph data. Any existing method can be used. For example, one data pair can be input into the generator at a time to train the generator once. Alternatively, multiple data pairs can be input into the generator at a time to train the generator once. Alternatively, all data pairs can be input into the generator at once to train the generator.

[0097] In some embodiments, the embodiments of the present application may train the generator in a batch manner, that is, the above S202 includes the following steps S202-A1 to S202-A6:

[0098] S202-A1. For each generator of the N sets of hyperparameters, obtain the i-th batch from the first training graph data according to the preset batch size.

[0099] The i-th batch includes at least one data pair, i is a positive integer from 1 to M, and M is the preset number of generator training times.

[0100] The embodiment of the present application does not limit the method of obtaining the batch from the first training graph data.

[0101] In one example, according to a preset batch size, batch size data pairs are randomly selected from the R data pairs included in the first training graph data each time as a batch, wherein different batches may contain the same data pairs.

[0102] In one example, according to a preset batch size, each R data pair included in the first training graph data is evenly divided into several parts, and each part is regarded as a batch.

[0103] In an embodiment of the present application, the training process of the generator under each of the N groups of hyper-parameters is the same. For the sake of convenience of description, this embodiment takes the training process of the generator under one group of hyper-parameters as an example, and the training process of the generator under other groups of hyper-parameters can be used as a reference.

[0104] S202-A2. Use the i-th batch to train the generator under this set of hyperparameters, so that the generator learns the feature probability distribution of the neighboring nodes of each central node under the given feature conditions of each central node in the i-th batch, and obtains the generator after i-th batch training.

[0105] S202-A3: Use the generator trained with the i-th batch to perform sampling to obtain the third sampling sample corresponding to the i-th batch output by the generator.

[0106] The embodiment of the present application uses a substantially consistent process for training the generator using each batch, and the training process of the generator is an iterative process. For example, after the generator is trained using the current batch, the parameters of the generator corresponding to the current batch are fixed, and the parameters of the generator fixed by the current batch are used as the initial parameters of the generator for the next batch training, and the iterations are performed in sequence until the generator training end condition is directly met. The generator training end condition may be that the generator loss reaches a preset value, or the generator training times reach a preset number of times.

[0107] Optionally, in this embodiment, the training termination condition of the generator is that the generator has been trained M times.

[0108] In an embodiment of the present application, after using each batch to train the generator, the quality of the generator after batch training is evaluated, and the generator with the best quality is determined as the target generator.

[0109] Among them, the method for evaluating the quality of the generator after batch training can be to use the trained generator for sampling, use the sampled samples obtained by sampling to train the second prediction model, and determine the quality of the generator after batch training based on the prediction results of the trained second prediction model.

[0110] The embodiments of the present application do not limit the specific network structure of the generator. For example, it can be many classic generation models such as Generative Adversarial Networks (GAN), Energy-Based Model (EBM), etc.

[0111] In one possible implementation, the generator is a generator in an adversarial network, which includes a generator and a discriminator. In this case, the above S202-A2 includes: inputting the i-th batch into the generator, the generator inputting the output result into the discriminator, the discriminator judging whether the sample output by the generator is true, and then reversely adjusting the parameters in the generator to obtain the generator after the i-th batch training. Correspondingly, the above S202-A3 includes: selecting K nodes from the first training graph data, randomly extracting a perturbation for each of the K nodes, inputting the first feature information of the node and the perturbation of the node into the generator after the i-th batch training, and obtaining the third sampling sample output by the generator after the i-th batch training. The number of the third sampling samples can be one or more, which can be specifically set.

[0112] In a possible implementation, the generator of the embodiment of the present application is a conditional variational autoencoder. Figure 4 As shown, the conditional variational autoencoder (VAE) is a directional graph generation model, including an encoding module and a decoding module. In some embodiments, the encoding module in the conditional variational autoencoder is called an encoder, and the decoding module is called a decoder. Figure 4 Based on the conditional variational autoencoder shown in FIG. 1 , the above S202-A2 includes the following steps S202-A21 to S202-A23:

[0113] S202-A21. For each data pair in the i-th batch, input the data pair into the encoding module to obtain a first variance and a first mean output by the encoding module.

[0114] S202-A22. Convert the first variance and the first mean into a first latent variable through a reparameterization trick, and input the first latent variable and the first feature information of the central node in the data pair into a decoding module to obtain an augmented feature vector output by the decoding module.

[0115] S202-A23: Use the augmented feature vector to train the encoding module and the decoding module to obtain the encoding module and the decoding module after the i-th batch training.

[0116] For the convenience of description, in the embodiment of the present application, the variance and mean output by the encoding module during the generator training process are respectively recorded as the first variance and the first mean.

[0117] Optionally, the first variance and the first mean conform to a Gaussian distribution.

[0118] The number of augmented feature vectors output by the decoding module can be one or more, and the specific number can be set. If there are multiple augmented feature vectors, the multiple augmented feature vectors are the same.

[0119] The augmented feature vector output by the decoding module is compared with the first feature information of the central node of the input encoding module. According to the difference between the two, the conditional variational autoencoder is adjusted to obtain the encoding module and decoding module after the i-th batch training.

[0120] In some embodiments, the loss function (or objective function) of the conditional variational autoencoder is as shown in formula (2):

[0121]

[0122]

[0123] in, is the loss function (or objective function) of the conditional variational autoencoder, and the encoding module is The decoding module is {f, g} are approximated by a neural network. are the parameters of the conditional variational autoencoder, represents the variational parameter, ψ represents the generating parameter, It is a normal distribution, with mathematical expectation μ and variance σ2. KL is a divergence operation, which is used to compare the closeness of two probability distributions.

[0124] The derivation process of the above formula (2) is introduced below:

[0125] In some embodiments, a conditional variational autoencoder is used to learn the latent variable distribution z j and the conditional distribution p(·|X i , z j ).

[0126] In order to derive the optimization problem of the conditional variational autoencoder, the logp in the above formula (1) is replaced by ψ (X j |Xi ) is written as the following formula (3):

[0127]

[0128] The Evidence Lower Bound (ELBO) in the above formula (3) is It can be written as the above formula (2), and then the loss function (or objective function) of the conditional variational autoencoder is derived.

[0129] According to the above, the encoding module and decoding module in the conditional variational autoencoder are trained using the i-th batch to obtain the encoding module and decoding module after the i-th batch training.

[0130] Next, the variational autoencoder after batch i training is sampled, specifically the decoding module after batch i training is sampled. Correspondingly, the sampling process in S202-A3 includes the following steps S202-A31 to S202-A33:

[0131] S202-A31. Select the first feature information of K nodes, where K is a positive integer.

[0132] Optionally, K nodes are randomly selected from the first training graph data.

[0133] Optionally, select K nodes from other graph data.

[0134] S202-A32: Randomly select a second variance and a second mean for each of the K nodes.

[0135] Optionally, the second variance and the second mean conform to a Gaussian distribution.

[0136] S202-A33. For each of the K nodes, the second variance and the second mean corresponding to the node are converted into a second latent variable through the reparameterization technique, and the second latent variable and the first feature information of the node are input into the decoding module after the i-th batch training to obtain the third sampling sample output by the decoding module.

[0137] After the training of the conditional variational autoencoder is completed, in the actual sampling process, the decoding module in the conditional variational autoencoder is used for sampling. During the training process, the encoding module helps the decoding module learn the feature probability distribution of the neighboring nodes around the central node.

[0138] The embodiment of the present application does not limit the sampling method using the generator trained with the i-th batch, and any existing sampling method can be used for sampling.

[0139] Optionally, an importance sampling method can be used to sample the generator after training the i-th batch to obtain the third sampling sample corresponding to the i-th batch output by the generator.

[0140] S202-A4: Input the third sample corresponding to the i-th batch into the preset second prediction model to obtain the prediction result corresponding to the i-th batch output by the second prediction model.

[0141] Figure 5 This is a schematic diagram of a training process involved in an embodiment of the present application, such as Figure 5 As shown in FIG, the generator is trained using the i-th batch to obtain the generator trained with the i-th batch, the generator trained with the i-th batch is used for sampling, the third sampling sample corresponding to the i-th batch is output, the third sampling sample corresponding to the i-th batch is input into the second prediction model, and the prediction result corresponding to the i-th batch output by the second prediction model is obtained.

[0142] In some embodiments, the second prediction model is a trained prediction model.

[0143] In some embodiments, the second prediction model is a prediction model that has been pre-trained with a preset number of steps.

[0144] The embodiment of the present application does not limit the specific type of the second prediction model.

[0145] In one example, the second prediction model is any one of graph neural networks such as a graph convolutional model (GCN), a graph attention network model (GAT), and a gated graph neural network model (GGNN).

[0146] In some embodiments, the second prediction model is a graph convolution model, and the prediction result of the graph convolution model is the classification result of the node.

[0147] According to the above steps, the generator is trained using the i-th batch to obtain the generator trained with the i-th batch, and the generator trained with the i-th batch is used for sampling to obtain the third sampling sample corresponding to the i-th batch. The third sampling sample is input into the second prediction model to obtain the prediction result of the second prediction model, and then the following steps S202-A5 and S202-A6 are executed.

[0148] S202-A5. Determine the uncertainty score corresponding to the i-th batch based on the prediction result corresponding to the i-th batch output by the second prediction model.

[0149] In some embodiments, the present application introduces importance sampling to capture the appropriate generator matrix to improve efficiency and accelerate the optimization of maximum likelihood estimation. In importance sampling, the present application introduces uncertainty scores to extract the most important samples.

[0150] For example, according to the preset uncertainty score formula, the uncertainty score corresponding to the i-th batch is determined.

[0151] In one possible implementation, the prediction result corresponding to the i-th batch output by the second prediction model is input into the following formula (4) to determine the uncertainty score corresponding to the i-th batch:

[0152]

[0153] in, For uncertain scores, is the sample sampled by the generator, T is the number of samples of Monte Carlo sampling, ω t are the parameters of the network for the tth Monte Carlo sampling. A high uncertainty score indicates that the network has very high uncertainty about the generated feature matrix and is therefore selected to improve the performance of the second prediction model.

[0154] The embodiment of the present application adopts importance sampling technology, which can better train the generator and enable the generator to provide more and richer samples to the downstream graph neural network.

[0155] In some embodiments, the second prediction model is a pre-trained model, and the embodiments of the present application do not involve the training process of the second prediction model.

[0156] In some embodiments, the above-mentioned second prediction model is trained together with the generator, that is, in an embodiment of the present application, the second prediction model and the generator are trained alternately.

[0157] When the second prediction model and the generator are trained alternately, the embodiment of the present application further includes a training process for the second prediction model after executing S202-A5, wherein the training process for the second prediction model includes the following steps:

[0158] S202-A5-1. Determine whether the uncertainty score corresponding to the i-th batch is the highest uncertainty score among the uncertainty scores corresponding to the previous i batches.

[0159] S202-A5-2. If it is determined that the uncertainty score corresponding to the i-th batch is the highest uncertainty score among the uncertainty scores corresponding to the previous i batches, then train the second prediction model a first number of times using the third sample output by the generator trained on the i-th batch. The second prediction model trained the first number of times is used to predict the prediction result of the third sample output by the generator trained on the i+1-th batch.

[0160] If it is determined that the uncertainty score corresponding to the i-th batch is not the highest uncertainty score among the uncertainty scores corresponding to the previous i batches, the current second prediction model is not trained.

[0161] In some embodiments, before training the variational generator, the second prediction model is pre-trained a second number of times. That is, in this embodiment of the present application, if i=1, before inputting the third sample corresponding to the i-th batch into the preset second prediction model and obtaining the prediction result corresponding to the i-th batch output by the second prediction model, that is, before S202-A4, the method of this embodiment of the present application further includes: pre-training the second prediction model a second number of times. Correspondingly, the above S202-A4 includes: inputting the third sample corresponding to the first batch into the second prediction model that has been trained a second number of times, and obtaining the prediction result corresponding to the first batch output by the second prediction model.

[0162] The training process of the generator involved in the embodiment of the present application includes: first pre-training the second prediction model a second number of times, using the first batch to train the generator under the hyperparameter to obtain the generator trained with the first batch, then sampling the generator trained with the first batch to obtain a third sample corresponding to the first batch. Then, the third sample corresponding to the first batch is input into the second prediction model that has been trained a second number of times to obtain the prediction result corresponding to the first batch output by the second prediction model, and based on the prediction result corresponding to the first batch, the uncertainty score corresponding to the first batch is determined.

[0163] Next, the generator under the hyperparameter is trained using the second batch to obtain the generator trained with the second batch. Then, the generator trained with the second batch is sampled to obtain the third sampling sample corresponding to the second batch. The third sampling sample corresponding to the second batch is then input into the second prediction model after the second number of trainings to obtain the prediction result corresponding to the second batch output by the second prediction model, and the uncertainty score corresponding to the second batch is determined based on the prediction result corresponding to the second batch. Check whether the uncertainty score corresponding to the second batch is greater than the uncertainty score corresponding to the first batch. If the uncertainty score corresponding to the second batch is greater than the uncertainty score corresponding to the first batch, the third sampling sample output by the generator trained with the second batch is used to train the current second prediction model for the first number of times.

[0164] Next, the generator under the hyperparameter is trained using the third batch to obtain the generator trained with the third batch. Next, the generator trained with the third batch is sampled to obtain the third sampling samples corresponding to the third batch. The third sampling samples corresponding to the third batch are then input into the second prediction model after the first number of trainings to obtain the prediction results corresponding to the third batch output by the second prediction model, and the uncertainty score corresponding to the third batch is determined based on the prediction results corresponding to the third batch. Check whether the uncertainty score corresponding to the third batch is greater than the uncertainty score corresponding to the first batch and greater than the uncertainty score corresponding to the second batch. If the uncertainty score corresponding to the third batch is greater than the uncertainty score corresponding to the first batch and greater than the uncertainty score corresponding to the second batch, then use the third sampling samples output by the generator trained with the third batch to train the current second prediction model again for the first number of times. Similarly, the training process of the generator and the training process of the second prediction model are iteratively performed until i=M.

[0165] S202-A6. Determine the generator trained with the batch corresponding to the highest uncertainty score among the uncertainty scores corresponding to the M batches as the generator trained under this set of hyperparameters.

[0166] The above steps S202-A1 to S202-A5 describe the process of determining the uncertainty score corresponding to the generator trained on the i-th batch, using the i-th batch as an example. The process of determining the uncertainty score corresponding to the generator trained on M batches is consistent with the above process of determining the uncertainty score corresponding to the generator trained on the i-th batch, and is not further described here with reference to the description of the above embodiment.

[0167] According to the above steps, after determining the uncertainty score corresponding to the generator trained for each batch in the M batches, the generator trained for the batch corresponding to the highest uncertainty score among the uncertainty scores corresponding to the M batches is determined as the generator trained under this set of hyperparameters.

[0168] The above embodiment is illustrated by taking the generator trained under one set of hyperparameters as an example. The process of determining the generator trained under each set of hyperparameters in N sets of hyperparameters is the same, and the above description can be referred to.

[0169] S203. Determine a target generator from the generators trained under N groups of hyperparameters, where the target generator is used to generate a second sampling sample, and the second sampling sample is used to train a preset graph neural network model.

[0170] In the embodiment of the present application, methods for determining the target generator from the generators trained under N sets of hyperparameters include but are not limited to the following methods:

[0171] Method 1: Use test samples to verify the accuracy of the generator trained under each set of hyperparameters, and determine the generator trained with the highest set of hyperparameters as the target generator.

[0172] Method 2: Use test samples to verify the prediction speed of the generator trained under each set of hyperparameters, and determine the generator trained under the set of hyperparameters with the fastest prediction speed as the target generator.

[0173] Method three: sample a group of generators trained with hyperparameters, use the sampled samples to train the first prediction model, and determine the target generator from the generators trained with N groups of hyperparameters based on the training effect of the first prediction model.

[0174] In the third method, the above S203 specifically includes the following steps S203-A1 and S203-A2:

[0175] S203-A1. Use the generator trained under this set of hyperparameters to perform sampling to obtain the first sampling sample output by the generator trained under this set of hyperparameters, and use the first sampling sample output by the generator trained under this set of hyperparameters to train the preset first prediction model to obtain the trained first prediction model corresponding to this set of hyperparameters.

[0176] S203-A2. Determine the prediction accuracy of the trained first prediction model corresponding to each set of hyperparameters.

[0177] S203-A3. Determine the generator trained under a set of hyperparameters corresponding to the first prediction model with the highest prediction accuracy as the target generator.

[0178] The graph neural network model can be viewed as a classification function for predicting graph node labels, which can be represented by the graph neural network classification estimator P θ (Y|A, X) is used to characterize the conditional distribution of label Y corresponding to the graph structure A and the feature matrix X, where θ is a parameter. Given the training sample {A, X, Y}, the parameter θ can be estimated using maximum likelihood estimation (MLE) according to the following formula (5):

[0179]

[0180] where K is the index of the nodes in the training set whose labels were seen during semi-supervised training.

[0181] In the embodiment of this application, Figure 6 As shown, the generator is used to generate sampling samples, and the sampling samples are used to train the first prediction model. Therefore, the embodiment of the present application proposes a new model in The sampled samples obtained by sampling the generator can be understood as the feature matrix generated by the generator for feature augmentation. Figure 6 The network model shown, the maximum likelihood estimation in the sample Based on this, we optimize a marginal distribution P, as shown in formula (6):

[0182]

[0183] Due to Bayesian tractability, P in formula (6) is decomposed into the product of two posterior probabilities, as shown in the following formula (7):

[0184]

[0185] in, and They are respectively represented as the probability distribution of the first prediction model and the generator approximation, θ is the parameter of the generator, and Φ is the parameter of the first prediction model. In the embodiment of the present application, once Q φ After being trained, you can φ sampling To optimize

[0186] Based on the above description, the process of determining the hyperparameters of the generator in the embodiment of the present application can be, for each set of N sets of hyperparameters trained under the generator, using the generator trained under the set of hyperparameters to sample, and obtain the first sampled sample output by the generator trained under the set of hyperparameters, wherein the process of extracting the first sampled sample is basically the same as the process of sampling to obtain the third sampled sample. For example, the first feature information of several nodes in the training graph data is selected, and a variance and mean are randomly selected for each node. For each node, the variance and mean corresponding to the node are converted into latent variables through the reparameterization technique, and the latent variable and the first feature information of the node are input into the generator trained under the set of hyperparameters to obtain the first sampled sample output by the generator trained under the set of hyperparameters. Then, the first sampled sample output by the generator trained under the set of hyperparameters is used to train the preset first prediction model to obtain the first prediction model corresponding to the trained set of hyperparameters. Referring to the above method, the first prediction model corresponding to each set of hyperparameters under the N sets of hyperparameters can be obtained. Then, the prediction accuracy of the first prediction model corresponding to the trained set of hyperparameters is determined respectively. For example, using test samples, the prediction accuracy of the trained first prediction model corresponding to each set of hyperparameters is tested, and the generator trained under a set of hyperparameters corresponding to the first prediction model with the highest prediction accuracy is determined as the target generator.

[0187] The embodiment of the present application does not limit the specific type of the first prediction model. For example, the first prediction model is a classification model or other prediction model.

[0188] In a possible implementation, the first prediction model is an image convolutional neural network model.

[0189] The model training method provided by the embodiment of the present application is obtained by obtaining first training graph data and N groups of hyperparameters, wherein the first training graph data includes R data pairs, each data pair consists of the first feature information of a central node in the graph network and the first feature information of a neighboring graph node of the central node, and N and R are both positive integers; the first training graph data is respectively input into the generator under each set of hyperparameters in the N groups of hyperparameters, so that the generator learns the feature probability distribution of the neighboring nodes of the central node under the given feature conditions of the central node, and obtains the generator trained under each set of hyperparameters in the N groups of hyperparameters; from the generator trained under the N groups of hyperparameters, a target generator is determined, and the target generator is used to generate a second sampling sample, and the second sampling sample is used to train a preset graph neural network model. That is, the generator of the embodiment of the present application uses the first feature of the central node on the graph as a condition to learn the feature probability distribution of the neighboring nodes near the central node, so that the generator not only learns the topological structure information of the graph, but also learns the feature information of the graph nodes. The generator is used for the problem of data augmentation, which solves the problem of overfitting of the model due to lack of sufficient samples during training, and greatly improves the robustness and generalization of the model.

[0190] The above describes the training process of the generator. The following describes the process of using the trained generator to assist in the training of other models.

[0191] Figure 7 A flow chart of a model training method provided in one embodiment of the present application is provided. Figure 8 A schematic diagram of model training provided in an embodiment of the present application is shown in FIG. Figure 7 and Figure 8 As shown, the model training method includes:

[0192] S701: Use the target generator to perform sampling to obtain a second sampling sample output by the target generator.

[0193] In this step, the process of sampling to generate the second sample is consistent with the process of sampling to generate the third sample, and the description of the above embodiment can be referred to.

[0194] In some embodiments, the generator is a conditional variational autoencoder, such as Figure 4 As shown, the conditional variational autoencoder includes an encoding module and a decoding module. During the training process, the encoding module is used to assist the decoding module in training. After the training is completed, during the sampling process, the decoding module is used to perform sampling. At this time, the above S701 includes the following steps:

[0195] S701-A1. Obtain second training graph data.

[0196] The second training graph data includes first feature information of a plurality of second nodes.

[0197] Among them, the original graph networks of the second training graph data and the first training graph data can be the same or different, or partially the same and partially different, and this embodiment of the application does not limit this.

[0198] S701-A2. Randomly select a third difference and a third mean for each of the plurality of second nodes.

[0199] In some embodiments, the third difference and the third mean are randomly selected arbitrary values.

[0200] In some embodiments, the third deviation and the third mean satisfy a Gaussian distribution.

[0201] S701-A3. For each second node among the multiple second nodes, the third difference and the third mean corresponding to the second node are converted into a third latent variable through the reparameterization technique, and the third latent variable and the first feature information of the second node are input into the target decoding module in the target generator to obtain P second sampling samples about the second node output by the target decoding module, where P is a positive integer.

[0202] The above P is a preset value and can be understood as a super parameter of the generator model.

[0203] The embodiment of the present application does not limit the specific value of P, which is determined based on actual conditions.

[0204] The above P second sampling samples are the same.

[0205] S702. Use the second sampling sample output by the target generator to train the preset graph neural network model.

[0206] The embodiments of the present application do not limit the specific type of graph neural network model, such as a graph convolutional network or a graph attention mechanism network.

[0207] The graph neural network model captures the interdependence of the graph through the message propagation mechanism between graph nodes, where the message propagation mechanism is expressed by the following formula (8):

[0208] H (l) =f(A,H l-1 )

[0209] Among them, f represents the special graph neural network layer in different models, H (l) is the hidden vector of layer l, and H 0 = X. For example, the representation of a graph convolutional network is in yes The degree matrix of and

[0210] In some embodiments, the above S702 includes inputting the second sampling sample output by the above target generator into the graph neural network model to train the graph neural network model.

[0211] In some embodiments, the above S702 includes S702-A: inputting the second training graph data and the second sampling sample output by the target decoding module into the graph neural network model to train the graph neural network model.

[0212] The embodiment of the present application does not limit the method of using the second training graph data and the second sampling samples in S702-A to train the graph neural network model.

[0213] In one possible implementation, S702-A includes: inputting the second training graph data and the second sampling sample into the feature extraction layer of the graph neural network model to obtain the second feature information of each second node in the second training graph data and the second feature information of the second sampling sample. Calculating the average value of the second feature information of the second node and the second feature information of the second sampling sample, and inputting the average value into the next graph neural network layer in the graph neural network model to obtain the output result of the graph neural network model. Adjusting the parameters in the graph neural network model according to the output result of the graph neural network model to obtain a trained graph neural network model.

[0214] In some embodiments, as Figure 9 As shown, the graph neural network model includes a first feature extraction unit, P second feature extraction units, and a prediction unit. In this case, the above S702-A includes:

[0215] S702-A1. For each second node among the plurality of second nodes, input the first feature information of the second node into a first feature extraction unit, and obtain the second feature information of the first node output by the first feature extraction unit.

[0216] S702-A2: Input the P second sampling samples of the second node into P second feature extraction units respectively to obtain second feature information of each of the P second sampling samples.

[0217] S702-A3: Concatenate the second feature information of the second node and the second feature information of the P second sampling samples to obtain concatenated second feature information.

[0218] S702-A4: Input the concatenated second feature information into the prediction unit to train the graph neural network model. For example, the concatenated second feature information is input into the prediction unit, which outputs a prediction result of the graph neural network model based on the concatenated second feature information. The parameters of the graph neural network model are adjusted based on the output result of the graph neural network model to obtain a trained graph neural network model.

[0219] It should be noted that Figure 9 The first feature extraction unit, the second feature extraction unit and the prediction unit are an example. The embodiment of the present application does not limit the specific network structure of the first feature extraction unit, the second feature extraction unit and the prediction unit, which is determined according to the specific network structure of the graph neural network model.

[0220] In some embodiments, if the graph neural network model is a graph convolutional model, then Figure 9 H in (2) Calculated according to the following formula (9):

[0221]

[0222] Where, Ⅱ represents the cascade operation, W is the network parameter of the first feature extraction unit, W0 to W P-1 are the network parameters of P second feature extraction units respectively.

[0223] This embodiment of the application uses the second sample output by the target generator as a training sample for the graph neural network model to solve the overfitting problem caused by small-label semi-supervised learning in graph deep learning. The data augmentation technology used in this embodiment of the application can be extended to any method used for small-label graphs.

[0224] The model training method of the embodiment of the present application can be applied to different graph neural network models, and can improve the training accuracy of the model based on the original neural network model, thereby improving the performance of the model.

[0225] The following experimental data further illustrates the technical effects of the model training method provided in the embodiments of the present application.

[0226] This application (abbreviated as LA-GNN) is compared with existing methods. Experiments are conducted on the Cora, Citeseer, and Pubmed data sets with its underlying graph learning algorithms GCN, GAT, GCNII, and GRAND. Each group of experiments is subjected to 100 random experiments. The experimental results are shown in Table 1:

[0227] Cora Citeseer Pubmed GCN 81.6％ 70.3％ 78.9％ GAT 83.0％ 70.4％ OOM GCNII 85.2％ 73.1％ 80.0％ GRAND 85.3％ 75.3％ 82.7％ LA-GCN 84.1％ 72.5％ 81.5％ LA-GAT 83.9％ 72.3％ OOM LA-GCNII 85.2％ 73.5％ 81.0％ LA-GRAND 85.8％ 75.8％ 83.2％

[0228] In this embodiment, the target generator is combined with GCN, GAT, GCNII, and GRAND to obtain the LA-GCN, LA-GAT, LA-GCNII, and LA-GRAND models, respectively. As shown in Table 1, the LA-GCN, LA-GAT, LA-GCNII, and LA-GRAND models of this embodiment have improved prediction accuracy on Cora, Citeseer, and Pubmed data compared to the existing GCN, GAT, GCNII, and GRAND.

[0229] The preferred embodiments of the present application are described in detail above in conjunction with the accompanying drawings. However, the present application is not limited to the specific details in the above embodiments. Within the technical concept of the present application, a variety of simple modifications can be made to the technical solution of the present application, and these simple modifications all fall within the scope of protection of the present application. For example, the various specific technical features described in the above specific embodiments can be combined in any suitable manner unless there is any contradiction. In order to avoid unnecessary repetition, the present application will not further explain various possible combinations. For another example, the various different embodiments of the present application can also be arbitrarily combined, and as long as they do not violate the ideas of the present application, they should also be regarded as the contents disclosed in the present application.

[0230] It should also be understood that in the various method embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0231] Combined with the above Figures 2 to 9 , describes the method embodiment of the present application in detail, and the following is combined with Figures 10 and 11 , describe in detail the device embodiments of the present application.

[0232] Figure 10 1 is a schematic block diagram of a model training apparatus provided in an embodiment of the present application. The apparatus 10 may be a computing device or a part of a computing device.

[0233] like Figure 10 As shown, the model training device 10 may include:

[0234] An acquisition unit 11 is configured to acquire first training graph data and N sets of hyperparameters, wherein the first training graph data includes R data pairs, each data pair consisting of first feature information of a central node in a graph network and first feature information of a neighboring graph node of the central node, where N and R are both positive integers;

[0235] A training unit 12 is configured to input the first training graph data into a generator under each of the N sets of hyperparameters, so that the generator learns the feature probability distribution of the neighboring nodes of the central node under the given feature conditions of the central node, and obtain a generator trained under each of the N sets of hyperparameters;

[0236] The determination unit 13 is used to determine a target generator from the generators trained under N groups of hyperparameters, and the target generator is used to generate a second sampling sample, and the second sampling sample is used to train a preset graph neural network model.

[0237] In some embodiments, the determination unit 13 is specifically used to sample each group of hyperparameters in the N groups of hyperparameters using the generator trained under the group of hyperparameters to obtain a first sampling sample output by the generator trained under the group of hyperparameters, and use the first sampling sample output by the generator trained under the group of hyperparameters to train a preset first prediction model to obtain a trained first prediction model corresponding to the group of hyperparameters; determine the prediction accuracy of the trained first prediction model corresponding to each group of hyperparameters respectively; and determine the generator trained under a group of hyperparameters corresponding to the first prediction model with the highest prediction accuracy as the target generator.

[0238] In some embodiments, the training unit 12 is specifically used to obtain an i-th batch from the first training graph data for the generator under each group of hyperparameters in the N groups of hyperparameters according to a preset batch size, wherein the i-th batch includes at least one data pair, wherein i is a positive integer from 1 to M, and M is a preset number of training times of the generator; use the i-th batch to train the generator under the group of hyperparameters so that the generator learns the feature probability distribution of the neighboring nodes of each central node under the given feature conditions of each central node in the i-th batch, and obtains the generator after the i-th batch training; use The generator after training the i-th batch is sampled to obtain the third sampling sample corresponding to the i-th batch output by the generator; the third sampling sample corresponding to the i-th batch is input into the preset second prediction model to obtain the prediction result corresponding to the i-th batch output by the second prediction model; according to the prediction result corresponding to the i-th batch output by the second prediction model, the uncertainty score corresponding to the i-th batch is determined; the generator trained by the batch corresponding to the highest uncertainty score among the uncertainty scores corresponding to M batches is determined as the generator trained under this group of hyperparameters.

[0239] In some embodiments, the generator is a conditional variational autoencoder, which includes an encoding module and a decoding module. The training unit 12 is specifically used to input the data pair into the encoding module for each data pair in the i-th batch to obtain the first variance and the first mean output by the encoding module; convert the first variance and the first mean into a first latent variable through a reparameterization technique, and input the first latent variable and the first feature information of the central node in the data pair into the decoding module to obtain the augmented feature vector output by the decoding module; use the augmented feature vector to train the encoding module and the decoding module to obtain the encoding module and the decoding module after training the i-th batch.

[0240] In some embodiments, the training unit 12 is specifically used to select the first feature information of K nodes, where K is a positive integer; randomly select a second variance and a second mean for each of the K nodes; for each of the K nodes, convert the second variance and second mean corresponding to the node into a second latent variable through a reparameterization technique, and input the second latent variable and the first feature information of the node into the decoding module after the i-th batch training to obtain a third sampling sample output by the decoding module.

[0241] In some embodiments, the training unit 12 is further used to determine whether the uncertainty score corresponding to the i-th batch is the highest uncertainty score among the uncertainty scores corresponding to the previous i batches; if it is determined that the uncertainty score corresponding to the i-th batch is the highest uncertainty score among the uncertainty scores corresponding to the previous i batches, the third sampling sample output by the generator after training the i-th batch is used to train the second prediction model a first number of times; wherein, the second prediction model after training the first number of times is used to predict the prediction result of the third sampling sample output by the generator after training the i+1-th batch.

[0242] In some embodiments, if i=1, in some embodiments, the training unit 12 is also used to pre-train the second prediction model a second number of times; the third sampling sample corresponding to the first batch is input into the second prediction model after the second number of training times to obtain the prediction result corresponding to the first batch output by the second prediction model.

[0243] In some embodiments, the training unit 12 is further used to use the target generator to perform sampling to obtain a second sampling sample output by the target generator; and use the second sampling sample output by the target generator to train a preset graph neural network model.

[0244] In some embodiments, the generator is a conditional variational autoencoder, which includes a decoding module and a training unit 12, which is specifically used to obtain second training graph data, wherein the second training graph data includes first feature information of multiple second nodes; randomly select a third difference and a third mean for each of the multiple second nodes; for each second node in the multiple second nodes, convert the third difference and the third mean corresponding to the second node into a third latent variable through a reparameterization technique, and input the third latent variable and the first feature information of the second node into the target decoding module in the target generator to obtain P second sampling samples about the second node output by the target decoding module, where P is a positive integer.

[0245] In some embodiments, the training unit 12 is specifically used to input the second training graph data and the second sampling sample output by the target decoding module into the graph neural network model to train the graph neural network model.

[0246] In some embodiments, the graph neural network model includes a first feature extraction unit, P second feature extraction units and a prediction unit, and the training unit 12 is specifically used to input the first feature information of each second node in the multiple second nodes into the first feature extraction unit to obtain the second feature information of the first node output by the first feature extraction unit; input the P second sampling samples of the second node into the P second feature extraction units respectively to obtain the second feature information of each second sampling sample in the P second sampling samples; cascade the second feature information of the second node and the second feature information of the P second sampling samples to obtain the cascaded second feature information; input the cascaded second feature information into the prediction unit to train the graph neural network model.

[0247] It should be understood that the device embodiment and the method embodiment may correspond to each other, and similar descriptions may refer to the method embodiment. To avoid repetition, they will not be described here. Specifically, Figure 10 The device shown can execute the embodiments of the above method, and the aforementioned and other operations and / or functions of each module in the device are respectively for implementing the method embodiments corresponding to the computing device, which will not be repeated here for the sake of brevity.

[0248] The apparatus of the embodiment of the present application is described above from the perspective of functional modules in conjunction with the accompanying drawings. It should be understood that the functional module can be implemented in hardware form, can be implemented by instructions in software form, or can be implemented by a combination of hardware and software modules. Specifically, the steps of the method embodiment in the embodiment of the present application can be completed by the hardware integrated logic circuit and / or software form instructions in the processor, and the steps of the method disclosed in the embodiment of the present application can be directly embodied as being executed by a hardware decoding processor, or can be executed by a combination of hardware and software modules in the decoding processor. Optionally, the software module can be located in a mature storage medium in the art such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps in the above method embodiment in conjunction with its hardware.

[0249] Figure 11 It is a schematic block diagram of a computing device provided in an embodiment of the present application, which is used to execute the above method embodiment.

[0250] like Figure 11 As shown, the computing device 30 may include:

[0251] The memory 31 and the processor 32 are configured to store a computer program 33 and transmit the program code 33 to the processor 32. In other words, the processor 32 can call and run the computer program 33 from the memory 31 to implement the method in the embodiment of the present application.

[0252] For example, the processor 32 may be configured to execute the above method steps according to instructions in the computer program 33 .

[0253] In some embodiments of the present application, the processor 32 may include but is not limited to:

[0254] General-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic device, discrete hardware components, etc.

[0255] In some embodiments of the present application, the memory 31 includes but is not limited to:

[0256] Volatile memory and / or non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link DRAM (SLDRAM), and direct RAM bus random access memory (DR RAM).

[0257] In some embodiments of the present application, the computer program 33 may be divided into one or more modules, which are stored in the memory 31 and executed by the processor 32 to implement the method for recording a page provided by the present application. The one or more modules may be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program 33 in the computing device.

[0258] like Figure 11 As shown, the computing device 30 may further include:

[0259] The transceiver 34 may be connected to the processor 32 or the memory 31 .

[0260] The processor 32 may control the transceiver 34 to communicate with other devices. Specifically, the processor 32 may send information or data to other devices or receive information or data sent by other devices. The transceiver 34 may include a transmitter and a receiver. The transceiver 34 may further include one or more antennas.

[0261] It should be understood that the various components in the computing device 30 are connected via a bus system, wherein the bus system includes not only a data bus but also a power bus, a control bus and a status signal bus.

[0262] According to one aspect of the present application, a computer storage medium is provided, on which a computer program is stored. When the computer program is executed by a computer, the computer is enabled to perform the method of the above-mentioned method embodiment. Alternatively, the present application also provides a computer program product containing instructions. When the computer is executed by the instructions, the computer is enabled to perform the method of the above-mentioned method embodiment.

[0263] According to another aspect of the present application, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computing device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computing device to perform the method of the above method embodiment.

[0264] In other words, when implemented using software, it can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the process or function according to the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) method. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more available media integrated. The available medium can be a magnetic medium (e.g., a floppy disk, a hard disk, a tape), an optical medium (e.g., a digital video disc (DVD)), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0265] Those skilled in the art will appreciate that the modules and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0266] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the modules is merely a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0267] Modules described as separate components may or may not be physically separate, and components displayed as modules may or may not be physical modules, i.e., they may be located in one place or distributed across multiple network elements. Some or all of the modules may be selected based on actual needs to achieve the purpose of the present embodiment. For example, the functional modules in the various embodiments of the present application may be integrated into a processing module, or each module may exist physically separately, or two or more modules may be integrated into a single module.

[0268] The above content is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A model training method, characterized in that: Applied to the field of artificial intelligence pharmaceutical manufacturing, the method includes: Obtain first training graph data and N sets of hyperparameters, where the first training graph data includes R data pairs, each data pair consisting of first feature information of a central node in the graph network and first feature information of a neighboring node of the central node. Utilize the collected drug data to learn the structure and feature information near different functional groups and summarize information on the local structure of the drug molecules, where N and R are both positive integers. Inputting the first training graph data into the generator under each set of hyperparameters in the N sets of hyperparameters respectively, so that the generator learns the feature probability distribution of the neighboring nodes of the central node under the given feature conditions of the central node, and obtaining the generator trained under each set of hyperparameters in the N sets of hyperparameters; For each set of hyperparameters in the N sets of hyperparameters, sampling is performed using the generator trained under the set of hyperparameters to obtain a first sample output by the generator trained under the set of hyperparameters, and the first sample output by the generator trained under the set of hyperparameters is used to train a preset first prediction model to obtain a trained first prediction model corresponding to the set of hyperparameters; The prediction accuracy of the trained first prediction model corresponding to each set of hyperparameters is determined respectively, and the generator trained under a set of hyperparameters corresponding to the first prediction model with the highest prediction accuracy is determined as the target generator. The target generator is used to generate a second sampling sample, and the second sampling sample is used to train a preset graph neural network model.

2. The method according to claim 1, characterized in that The step of inputting the first training graph data into a generator under each set of hyperparameters in the N sets of hyperparameters, so that the generator learns the feature probability distribution of the neighboring nodes of the central node under the given feature conditions of the central node, and obtaining a generator trained under each set of hyperparameters in the N sets of hyperparameters, comprises: For the generator under each set of hyperparameters in the N sets of hyperparameters, obtain an i-th batch from the first training graph data according to a preset batch size, where the i-th batch includes at least one data pair, where i is a positive integer from 1 to M, and M is a preset number of generator training times; Using the i-th batch to train the generator under the set of hyperparameters, so that the generator learns the feature probability distribution of the neighboring nodes of each central node under the given feature conditions of each central node in the i-th batch, to obtain the generator after the i-th batch training; Perform sampling using the generator trained on the i-th batch to obtain a third sample corresponding to the i-th batch output by the generator; Inputting the third sample corresponding to the i-th batch into a preset second prediction model to obtain a prediction result corresponding to the i-th batch output by the second prediction model; Determining an uncertainty score corresponding to the i-th batch according to the prediction result corresponding to the i-th batch output by the second prediction model; The generator trained with the batch corresponding to the highest uncertainty score among the uncertainty scores corresponding to the M batches is determined as the generator trained under this set of hyperparameters.

3. The method according to claim 2, characterized in that The generator is a conditional variational autoencoder, which includes an encoding module and a decoding module. The generator under the set of hyperparameters is trained using the i-th batch to enable the generator to learn the feature probability distribution of the neighboring nodes of the central node under the given feature conditions of the central node in the i-th batch, and obtain the generator after the i-th batch training, including: For each data pair in the i-th batch, input the data pair into the encoding module to obtain a first variance and a first mean output by the encoding module; Converting the first variance and the first mean into a first latent variable through a reparameterization technique, and inputting the first latent variable and the first feature information of the central node in the data pair into the decoding module to obtain an augmented feature vector output by the decoding module; The encoding module and the decoding module are trained using the augmented feature vector to obtain the encoding module and the decoding module after the i-th batch training.

4. The method according to claim 3, characterized in that The step of sampling using the generator trained with the i-th batch to obtain a third sample output by the generator includes: Select the first feature information of K nodes, where K is a positive integer; Randomly select a second variance and a second mean for each of the K nodes; For each of the K nodes, the second variance and the second mean corresponding to the node are converted into a second latent variable through the reparameterization technique, and the second latent variable and the first feature information of the node are input into the decoding module after the i-th batch training to obtain a third sampling sample output by the decoding module.

5. The method according to claim 2, characterized in that The method further comprises: Determine whether the uncertainty score corresponding to the i-th batch is the highest uncertainty score among the uncertainty scores corresponding to the previous i batches; If it is determined that the uncertainty score corresponding to the i-th batch is the highest uncertainty score among the uncertainty scores corresponding to the previous i batches, training the second prediction model a first number of times using the third sampling sample output by the generator after training the i-th batch; Among them, the second prediction model after training the first number of times is used to predict the prediction result of the third sampling sample output by the generator after the i+1th batch training.

6. The method according to claim 5, characterized in that If i=1, before inputting the third sample corresponding to the i-th batch into a preset second prediction model and obtaining the prediction result corresponding to the i-th batch output by the second prediction model, the method further includes: pre-training the second prediction model a second number of times; Inputting the third sample corresponding to the i-th batch into a preset second prediction model to obtain a prediction result corresponding to the i-th batch output by the second prediction model includes: The third sampling sample corresponding to the first batch is input into the second prediction model after training for a second number of times to obtain the prediction result corresponding to the first batch output by the second prediction model.

7. The method according to any one of claims 1 to 6, characterized in that The method further comprises: Perform sampling using the target generator to obtain a second sample output by the target generator; The preset graph neural network model is trained using the second sampling sample output by the target generator.

8. The method according to claim 7, characterized in that The target generator is a target conditional variational autoencoder, and the target conditional variational autoencoder includes a target decoding module. The sampling using the target generator to obtain a second sample output by the target generator includes: Acquire second training graph data, where the second training graph data includes first feature information of a plurality of second nodes; Randomly selecting a third difference and a third mean for each of the plurality of second nodes; For each second node among the multiple second nodes, the third difference and the third mean corresponding to the second node are converted into a third latent variable through the reparameterization technique, and the third latent variable and the first feature information of the second node are input into the target decoding module in the target generator to obtain P second sampling samples about the second node output by the target decoding module, where P is a positive integer.

9. The method according to claim 8, characterized in that The training of a preset graph neural network model using the second sampling sample output by the target generator includes: The second training graph data and the second sampling sample output by the target decoding module are input into the graph neural network model to train the graph neural network model.

10. The method according to claim 9, characterized in that The graph neural network model includes a first feature extraction unit, P second feature extraction units, and a prediction unit. The second training graph data and the second sample output by the target decoding module are input into the graph neural network model to train the graph neural network model, including: For each second node among the plurality of second nodes, input the first feature information of the second node into the first feature extraction unit to obtain the second feature information of the second node output by the first feature extraction unit; Inputting the P second sampling samples of the second node into the P second feature extraction units respectively to obtain second feature information of each of the P second sampling samples; cascading the second feature information of the second node and the second feature information of the P second sampling samples to obtain cascaded second feature information; The cascaded second feature information is input into the prediction unit to train the graph neural network model.

11. A model training device, characterized in that: Applied in the field of artificial intelligence pharmaceutical manufacturing, the device comprises: an acquisition unit, configured to acquire first training graph data and N sets of hyperparameters, wherein the first training graph data includes R data pairs, each data pair consisting of first feature information of a central node in the graph network and first feature information of a neighboring node of the central node, and utilizing the collected drug data to learn structural and feature information near different functional groups and summarize information on the local structure of drug molecules, wherein N and R are both positive integers; a training unit, configured to input the first training graph data into a generator under each of the N sets of hyperparameters, so that the generator learns the feature probability distribution of the neighboring nodes of the central node under given feature conditions of the central node, and obtain a generator trained under each of the N sets of hyperparameters; A determination unit is used to sample, for each set of hyperparameters in the N sets of hyperparameters, using the generator trained under the set of hyperparameters to obtain a first sampling sample output by the generator trained under the set of hyperparameters, and to train a preset first prediction model using the first sampling sample output by the generator trained under the set of hyperparameters to obtain the trained first prediction model corresponding to the set of hyperparameters; determine the prediction accuracy of the trained first prediction model corresponding to each set of hyperparameters respectively, and determine the generator trained under a set of hyperparameters corresponding to the first prediction model with the highest prediction accuracy as a target generator, the target generator being used to generate a second sampling sample, and the second sampling sample being used to train a preset graph neural network model.

12. A computing device, characterized in that including processor and memory; The memory is used to store computer programs; The processor is configured to execute the computer program to implement the method according to any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that Used to store a computer program, wherein the computer program causes a computer to execute the method according to any one of claims 1 to 10.

Images