Generative adversarial learning knowledge graph completion method for large-scale complex device
Through the generative adversarial meta-learning method, combined with the conditional generation adversarial network and meta-learning architecture, the problem of data quantity and semantic information neglected in knowledge graph completion of large-scale complex devices is solved, efficient and accurate knowledge graph completion is achieved, and the interpretability of inference results is improved.
Patent Information
- Application Number
- CN202510070435.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-16
- Publication Date
- 2025-06-13
AI Technical Summary
When processing knowledge graphs of large-scale complex devices, the existing knowledge graphs are limited by the amount of data and the ignorance of complex semantic information, resulting in low completion accuracy and efficiency, affecting the effectiveness of downstream applications.
Generative adversarial meta-learning method is adopted to build a conditional generative adversarial network and meta-learning architecture, combining hierarchical knowledge graphs and text semantic background knowledge to achieve the completion of the knowledge graph, and through multi-task generation models and training strategies for local constraints and global gradients, the accuracy and computational efficiency of the completion are improved.
It realizes efficient completion of small sample knowledge graphs, improves the accuracy and computing efficiency of large-scale complex knowledge graph completion, and gives certain interpretability to the inference results.
Smart Images
Figure CN120146175A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the fields of data processing and knowledge completion, and relates to a method for knowledge graph completion, specifically to a generative adversarial meta-learning knowledge graph completion method for large-scale complex devices. Background Art
[0002] With the increasingly in-depth exploration of space by humans, it is urgent to conduct in-depth research on the space environment effects of spacecraft materials, devices, and their functional systems. The national major scientific and technological infrastructure of "Space Environment Ground Simulation Device", building an internationally leading space comprehensive environment ground simulation platform to achieve the simulation of major space environmental factors and their effects, is the key means to solve the above problems.
[0003] The space environment ground simulation device is a large-scale complex device with a variety of experimental equipment. In the large-scale complex domain knowledge graph constructed by it, there is a dataset with a long-tail distribution. Due to the incompleteness of knowledge and the existence of low-frequency knowledge samples, existing knowledge graph completion methods have problems such as being limited by the amount of data and ignoring complex semantic information, reducing the accuracy and efficiency of knowledge graph completion, and seriously affecting the effectiveness of the knowledge graph in downstream applications. Summary of the Invention
[0004] In order to complete the small-sample knowledge graph, improve the accuracy and computational efficiency of large-scale complex knowledge graph completion, and endow the inference result with a certain degree of interpretability, the present invention provides a generative adversarial meta-learning knowledge graph completion method for large-scale complex devices.
[0005] The purpose of the present invention is achieved through the following technical solutions:
[0006] A generative adversarial meta-learning knowledge graph completion method for large-scale complex devices includes the following steps:
[0007] Step S1, Preparatory Link:
[0008] Construct a hierarchical knowledge graph, and based on this, construct a conditional generative adversarial network and a meta-learning architecture. The specific steps are as follows:
[0009] Step S1.1, Construct a Conditional Generative Adversarial Network: For the hierarchical structure characteristics in large-scale complex devices, introduce text semantic background knowledge to construct a conditional generative adversarial network to constrain the completed knowledge. The specific steps are as follows:
[0010] Step S1.1.1, Analyze the hierarchical structure characteristics in large-scale complex devices and determine the number of hierarchical structures of the space environment ground simulation device;
[0011] Step S1.1.2: Obtain the conditional vector c of the conditional generative adversarial network according to the hierarchical structure characteristics;
[0012] Step S1.1.3: Obtain the entity vector and relationship vector constrained by the conditional vector, and construct the adjacency matrix A. The adjacency matrix is a two-dimensional matrix, where the rows represent entities, the columns represent relationships, and the elements in the matrix represent the connection situation between entities and relationships;
[0013] Step S1.1.4: Obtain the random noise vector z of the conditional generative adversarial network. The noise vector z selects the Gaussian distribution N(0, 1) as the source of noise, and performs random sampling on each dimension to obtain the values of each dimension of the noise vector z. Subsequently, the values sampled on each dimension are combined into a noise vector;
[0014] Step S1.1.5: Obtain the complete input of the generator of the conditional generative adversarial network. The complete input of the generator of the conditional generative adversarial network consists of three parts: the conditional vector c, the adjacency matrix A, and the noise vector z;
[0015] Step S1.1.6: Obtain the complete input of the discriminator of the conditional generative adversarial network. The complete input of the discriminator of the conditional generative adversarial network consists of three parts: the randomly sampled relationship vector, the relationship vector generated by the generator, and the conditional vector;
[0016] Step S1.2: Construct a meta-learning architecture: Perform multi-task settings for the hierarchical structure characteristics in large-scale complex devices, and construct a meta-learning architecture. Each task contains a support set and a query set, and shares a feature extractor or encoder to process the input data of different tasks. The specific steps are as follows:
[0017] Step S1.2.1: According to the number of hierarchical structures of the spatial environment ground simulation device determined in Step S1.1.1, perform multi-task settings on the meta-learning framework, and determine the support set and query set of each task, and share the feature extractor or encoder to process the input data of different tasks, where: The support set is a training data set used to learn general knowledge or adapt to new tasks, and the query set is used to evaluate the generalization ability and adaptability of the multi-task generation model;
[0018] Step S1.2.2: Construct the base learner of the meta-learning framework. The base learner consists of three parts: a generator set, a data aggregator, and a shared discriminator, where: The generator set consists of multiple task generators, and each task has a task generator for generating task-related content; The data aggregator is used to summarize the meta-data completion relationships sampled from multiple task subgraphs; The shared discriminator is a shared component. By sharing a discriminator, the discriminator can learn between multiple tasks;
[0019] Step S1.2.3, constructing the meta-learner of the meta-learning framework: constructing the meta-learner of the meta-learning framework through the meta-gradient formed by accumulating the objective function of the conditional generative adversarial network and the gradients of each task;
[0020] Step S2, embedding phase:
[0021] Call the conditional generative adversarial network and the meta-learning framework constructed in the previous phase, embed the conditional generative adversarial network into the base learner in the meta-learning framework, and embed the objective function of the conditional generative adversarial network into the meta-learner in the meta-learning framework to construct a multi-task generation model;
[0022] Step S3, optimization phase:
[0023] Independently optimize the base learner and update the meta-learner globally with gradients. Combine the training strategy of the multi-task generation model with local constraints and global gradients to adjust the multi-task generation model. The specific steps are as follows:
[0024] Step S3.1, independently optimize the base learner in the multi-task generation model;
[0025] Step S3.2, enable the meta-learner to achieve migration between multiple related inference tasks through global gradient update, so as to quickly adapt to new tasks;
[0026] Step S3.3, adjust the multi-task generation model through the multi-task generation model training strategy, decouple the optimization of the multi-task generation model, and adaptively adjust the objective function and the knowledge graph completion strategy. Among them: the multi-task generation model training strategy consists of two parts: local constraints and global gradients.
[0027] Compared with the prior art, the present invention has the following advantages:
[0028] 1. In view of the complex semantic information contained in the large-scale knowledge graph, the present invention introduces a conditional vector based on hierarchical background knowledge in the generative adversarial network to constrain the complementary knowledge, and combines it with the meta-learning framework to construct a multi-task generation model, which can complete the completion of the small-sample knowledge graph and endow the inference result with a certain degree of interpretability.
[0029] 2. The present invention designs a training strategy that combines local constraints and global gradients, independently optimizes the base learner in the multi-task generation model, and the meta-learner performs joint learning through gradient update to achieve migration between multiple related inference tasks, and adaptively adjusts the objective function and the completion strategy, which can improve the accuracy and computational efficiency of large-scale complex knowledge graph completion. Description of the Drawings
[0030] Figure 1It is a block diagram of a generative adversarial meta-learning knowledge graph completion method for large-scale complex devices.
[0031] Figure 2 It is a flowchart of a generative adversarial meta-learning knowledge graph completion method for large-scale complex devices.
[0032] Figure 3 It is a specific flowchart of the generative adversarial meta-learning knowledge graph completion method for large-scale complex devices in the embodiment. Detailed implementation manners
[0033] The technical solutions of the present invention will be further described below in conjunction with the accompanying drawings, but are not limited thereto. Any modification or equivalent replacement of the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention shall be covered by the protection scope of the present invention.
[0034] The present invention provides a generative adversarial meta-learning knowledge graph completion method for large-scale complex devices. As Figure 1 shown, the method includes three major links: a preprocessing link, an embedding link, and an optimization link. Among them: the preprocessing link includes the construction of a hierarchical knowledge graph, and based on this, a conditional generative adversarial network and a meta-learning architecture are constructed; the embedding link includes the construction of a base learner and a meta-learner in the meta-learning architecture, and based on this, a multi-task generation model is constructed; the optimization link includes the independent optimization of the base learner and the global gradient update of the meta-learner; the embedding link calls the conditional generative adversarial network and the meta-learning framework constructed in the preprocessing link, and embeds the conditional generative adversarial network into the meta-learning process; the optimization link combines the training strategies of local constraints and global gradients to adjust the multi-task generation model. As Figure 2 shown, it specifically includes the following steps:
[0035] S1: Through the preprocessing link, aiming at the hierarchical structure characteristics in large-scale complex devices, text semantic background knowledge is introduced to construct a conditional generative adversarial network to constrain the complementary knowledge;
[0036] S2: Through the preprocessing link, a meta-learning architecture is constructed by setting multiple tasks aiming at the hierarchical structure characteristics in large-scale complex devices. Each task includes a support set and a query set, and a feature extractor or encoder is shared to process the input data of different tasks;
[0037] S3: Through the embedding link, the conditional generative adversarial network and the meta-learning framework constructed in the preprocessing link are called, the conditional generative adversarial network is embedded into the base learner in the meta-learning framework, and the objective function of the conditional generative adversarial network is embedded into the meta-learner in the meta-learning framework to construct a multi-task generation model;
[0038] S4: Through the optimization process, the base learners in the multi-task generation model are independently optimized, and the meta-learner realizes the transfer between multiple related inference tasks through global gradient update, so as to quickly adapt to new tasks.
[0039] S5: The multi-task generation model is adjusted through the training strategy of the multi-task generation model in the optimization process, and the objective function and knowledge graph completion strategy are adaptively adjusted.
[0040] In the present invention, large-scale complex devices should have typical hierarchical structure characteristics, including but not limited to system-subsystem-unit hierarchy, cyber-physical asset hierarchy, business logic hierarchy, etc.
[0041] In the present invention, the conditional generative adversarial network uses the hierarchical information generated from the text semantic background knowledge as the constraint condition of the generative adversarial network; the meta-learning framework selects the optimization-based model-agnostic meta-learning framework Model-Agnostic Meta-Learning (MAML), which is compatible with any model using the gradient descent algorithm.
[0042] In the present invention, the complete input of the conditional generative adversarial network generator consists of three parts: the conditional vector c, the adjacency matrix A, and the noise vector z. Among them: the adjacency matrix A is composed of the entity vector and the relationship vector constrained by the conditional vector; the noise vector z selects the Gaussian distribution N(0,1) as the source of noise, and random sampling is performed on each dimension to obtain the values of each dimension of the noise vector z, and then the values sampled on each dimension are combined into a noise vector.
[0043] In the present invention, the complete input of the conditional generative adversarial network discriminator consists of three parts: the randomly sampled relationship vector, the relationship vector generated by the generator, and the conditional vector.
[0044] In the present invention, the support set is the training data set used to learn general knowledge or adapt to new tasks, and the query set is used to evaluate the generalization ability and adaptability of the multi-task generation model.
[0045] In the present invention, the base learner of the meta-learning framework consists of three components: a generator set, a data aggregator, and a shared discriminator. Among them: the generator set is composed of multiple conditional generative adversarial network generators, and each task has a conditional generative adversarial network generator for generating task-related content; the data aggregator is used to summarize the metadata completion relationships sampled from multiple task subgraphs; the shared discriminator is a shared component, and by sharing a discriminator, the conditional generative adversarial network discriminator can learn between multiple tasks.
[0046] In the present invention, the multi-task generation model training strategy consists of two parts: local constraint and global gradient.
[0047] Example:
[0048] This example provides a method for generating adversarial meta-learning knowledge graph completion for large-scale complex devices, as Figure 3 shown, the method includes the following:
[0049] Step S1: Analyze the hierarchical structure characteristics in large-scale complex devices:
[0050] Analyze the basis for hierarchical division of the space environment ground simulation device and determine the number of hierarchical structures N of the space environment ground simulation device.
[0051] In this step, the basis for hierarchical division of the space environment ground simulation device selects system-subsystem-unit hierarchy, where: the system includes 7 experimental systems, namely the comprehensive environment simulation subsystem (referred to as the "comprehensive subsystem" for short), the space life science subsystem (referred to as the "life subsystem" for short), the device ion irradiation subsystem (referred to as the "device subsystem" for short), the micro-mechanism analysis subsystem (referred to as the "micro subsystem" for short), the ion accelerator subsystem, the space magnetic environment simulation and research system (referred to as the "magnetic system" for short), and the space plasma environment simulation and research system (referred to as the "plasma system" for short).
[0052] In this step, the number of hierarchical structures N of the space environment ground simulation device is N = 7.
[0053] Step S2: According to the hierarchical structure characteristics, obtain the conditional vector c of the conditional generative adversarial network.
[0054] In this step, the conditional vector of the conditional generative adversarial network is obtained by using one-hot encoding according to the number of hierarchical structures of the space environment ground simulation device.
[0055] Step S3: Obtain the entity vector and relationship vector constrained by the conditional vector, and construct the adjacency matrix A.
[0056] In this step, the adjacency matrix is a two-dimensional matrix, where the rows represent entities, the columns represent relationships, and the elements in the matrix represent the connection situation between entities and relationships.
[0057] Step S4: Obtain the random noise vector z of the conditional generative adversarial network.
[0058] In this step, the Gaussian distribution N(0,1) is selected as the source of noise, and random sampling is performed on each dimension to obtain the values of each dimension of the noise vector. Subsequently, the values sampled on each dimension are combined into a noise vector.
[0059] Step S5: Obtain the complete input of the generator of the conditional generative adversarial network.
[0060] In this step, the adjacency matrix is encoded into a vector of fixed length and concatenated with the generated noise vector to form the complete input of the conditional generative adversarial network generator.
[0061] Step S6: Obtain the complete input of the conditional generative adversarial network discriminator.
[0062] In this step, the real relationship and the generated relationship are mapped into vector representations and concatenated with the embedding vector of the conditional information to form the feature vector of the input of the conditional generative adversarial network discriminator.
[0063] In this step, the real relationship is a relationship randomly sampled from the hierarchical knowledge graph data and used as a benchmark for training the discriminator; the generated relationship is a synthetic relationship sample generated by the conditional generative adversarial network generator according to the defined conditions.
[0064] Step S7: Perform multi-task setting on the meta-learning framework and determine the support set and query set for each task.
[0065] In this step, according to the number of hierarchical structures of the space environment ground simulation device determined in Step S1, perform multi-task setting. Each task includes a support set and a query set and shares a feature extractor or encoder to process the input data of different tasks.
[0066] In this step, the support set is a training data set used to learn general knowledge or adapt to new tasks, and the query set is used to evaluate the generalization ability and adaptability of the model.
[0067] Step S8: Construct the base learner of the meta-learning framework.
[0068] In this step, the base learner of the meta-learning framework is constructed through the collaborative action of components such as a generator set, a data aggregator, and a shared discriminator, mainly including the following steps:
[0069] Step S81: According to the differences in the graph structures and features of different tasks, decompose the large-scale knowledge graph into multiple task subgraphs. Each task subgraph is regarded as an independent task, corresponding to the aforementioned conditional vector c, and its corresponding conditional generative adversarial network generator is responsible for generating task-related content.
[0070] Step S82: Design a conditional generative adversarial network generator for each task subgraph to form a generator set. Each conditional generative adversarial network generator generates entity and relationship vectors related to the task by receiving the conditional input of the corresponding task subgraph.
[0071] Step S83: Introduce a data aggregator to aggregate the meta-data completion relationships sampled from multiple task subgraphs. The aggregated data is passed to the conditional generative adversarial network discriminator to evaluate the authenticity of the generated content.
[0072] Step S84: Introduce a shared discriminator component to achieve learning sharing among multiple tasks through the shared discriminator, thereby improving the performance and robustness of the discriminator of the conditional generative adversarial network.
[0073] Step S9: Construct a meta-learner for the meta-learning framework.
[0074] In this step, a meta-learner for the meta-learning framework is constructed through the meta-gradient formed by accumulating the objective function of the conditional generative adversarial network and the gradients of each task.
[0075] Step S10: Independently optimize the base learner in the model.
[0076] In this step, in a specific task, the base learner is trained using limited known information. The conditional generative adversarial network generator and the conditional generative adversarial network discriminator corresponding to the specific task update their loss functions through alternating optimization, where: when updating the conditional generative adversarial network discriminator, its performance is improved by maximizing the objective function; while when updating the conditional generative adversarial network generator, the performance optimization of the conditional generative adversarial network generator is achieved by minimizing the objective function.
[0077] Step S11: Enable the meta-learner to achieve transfer between multiple related inference tasks through global gradient update.
[0078] In this step, the backpropagation algorithm is adopted to calculate the gradient of the generator parameters using the generator loss function in each task.
[0079] In this step, calculate the gradient of the generator loss function with respect to the initial parameters of the meta-learning framework, that is, the calculation of the meta-gradient is based on the gradients of multiple tasks and is accumulated.
[0080] In this step, the initial parameters of the meta-learning framework are updated by using the meta-gradient to achieve global gradient update, enabling the meta-learner to achieve effective transfer learning between multiple related inference tasks.
[0081] Step S12: Decouple the optimization of the multi-task generation model and adaptively adjust the objective function and the knowledge graph completion strategy.
[0082] In this step, through an alternating training process, the optimization of the objective function of the base learner realizes the adaptability of the multi-task generation model, enabling the generative adversarial network to more effectively generate real data that meets the task requirements under specific tasks.
[0083] In this step, in the meta-learning stage, the meta-gradient represents the parameter update rule shared by multiple tasks, enabling the multi-task generation model to quickly adapt to unseen tasks.
[0084] Example:
[0085] The performance of the link prediction metric HITS@n evaluation method is evaluated on the constructed Space Environment Ground Simulation Device Equipment Dataset SESRI-Dev. The Space Environment Ground Simulation Device Equipment Dataset SESRI-Dev includes 10,000 entities, and the number of tasks in the SESRI-Dev dataset and the number of relationships used in each task are divided according to a three-level hierarchical structure of system / sub-system - sub-system - unit. The number of tasks is divided into 8, 20, and 50, and the number of relationships used in each task is divided into three types, namely 50, 100, and 200. The link prediction metric HITS@n refers to the average proportion of triples with a rank less than n in link prediction. In this example, n = 10 is selected, and its expression is as follows:
[0086]
[0087] In the formula, S is the set of triples, |S| is the number of triple sets, rank i refers to the link prediction rank of the i-th triple, and Π is the indicator function (the function value is 1 if the condition holds; otherwise, it is 0).
[0088] The experimental results are shown in Table 1:
[0089] Table 1
[0090]
[0091] In terms of improving model performance, there is a certain complementary relationship between the number of tasks and the number of relationships in a task. When the number of tasks is limited, increasing the number of relationships helps to make up for the disadvantage of insufficient task numbers and enriches the learning content of the model in a single task. Similarly, when the number of relationships is small, by increasing the number of tasks to improve the knowledge transfer ability, performance improvement can also be achieved.
[0092] After the number of tasks reaches a certain scale, the marginal benefit of simply increasing the number of relationships decreases. This phenomenon may be related to the increased complexity of the model. When both the number of tasks and the number of relationships are high, the model may have difficulty efficiently processing the data noise brought by too many tasks.
Claims
1. A generative adversarial meta-learning knowledge graph completion method for large-scale complex devices, characterized by The method comprises the following steps: Step S1, pre-stage: Build a hierarchical knowledge graph and use it as a support to build conditional generative adversarial networks and meta-learning architectures; Step S2: Embedding: Call the conditional generative adversarial network and meta-learning framework constructed in the previous stage, embed the conditional generative adversarial network into the base learner in the meta-learning framework, embed the objective function of the conditional generative adversarial network into the meta-learner in the meta-learning framework, and build a multi-task generation model; Step S3, optimization phase: Independently optimize the base learner, globally update the meta-learner with gradients, and combine the multi-task generative model training strategy with local constraints and global gradients to adjust the multi-task generative model.
2. The generative adversarial meta-learning knowledge graph completion method for large-scale complex devices according to claim 1 is characterized in that The specific steps of step S1 are as follows: Step S1.1, constructing a conditional generative adversarial network: In view of the hierarchical characteristics of large-scale complex devices, text semantic background knowledge is introduced to construct a conditional generative adversarial network to constrain the completion knowledge; Step S1.2, construct a meta-learning architecture: A multi-task setting is performed for the hierarchical characteristics in large-scale complex devices, and a meta-learning architecture is constructed. Each task contains a support set and a query set, and shares a feature extractor or encoder to process the input data of different tasks.
3. The generative adversarial meta-learning knowledge graph completion method for large-scale complex devices according to claim 2 is characterized in that The specific steps of step S1.1 are as follows: Step S1.1.1, analyzing the hierarchical characteristics in large-scale complex devices and determining the number of hierarchical structures of the space environment ground simulation device; Step S1.1.2, obtaining a conditional vector c of a conditional generative adversarial network according to the hierarchical structure characteristics; Step S1.1.3, obtain the entity vector and relationship vector constrained by the conditional vector, and construct an adjacency matrix A, wherein the adjacency matrix is a two-dimensional matrix, wherein rows represent entities, columns represent relationships, and elements in the matrix represent the connection between entities and relationships; Step S1.1.4, obtain a random noise vector z of the conditional generative adversarial network, wherein the noise vector z selects a Gaussian distribution N(0,1) as the source of noise, and performs random sampling on each dimension to obtain the values of each dimension of the noise vector z, and then combines the values sampled on each dimension into a noise vector; Step S1.1.5, obtaining a complete input of a conditional generative adversarial network generator, wherein the complete input of the conditional generative adversarial network generator consists of three parts: a conditional vector c, an adjacency matrix A, and a noise vector z; Step S1.1.6, obtain the complete input of the conditional generative adversarial network discriminator, wherein the complete input of the conditional generative adversarial network discriminator consists of three parts: a randomly sampled relationship vector, a relationship vector generated by the generator, and a conditional vector.
4. The generative adversarial meta-learning knowledge graph completion method for large-scale complex devices according to claim 3 is characterized in that The large-scale complex device has typical hierarchical characteristics, including but not limited to system-subsystem-unit hierarchy, information-physical asset hierarchy, and business logic hierarchy.
5. The generative adversarial meta-learning knowledge graph completion method for large-scale complex devices according to claim 2 is characterized in that The specific steps of step S1.2 are as follows: Step S1.2.1, according to the hierarchical structure number of the space environment ground simulation device determined in step S1.1.1, perform multi-task setting for the meta-learning framework, determine the support set and query set of each task, and share the feature extractor or encoder to process the input data of different tasks, wherein: the support set is a training data set used to learn general knowledge or adapt to new tasks, and the query set is used to evaluate the generalization ability and adaptability of the multi-task generation model; Step S1.2.2, construct a base learner of the meta-learning framework, the base learner consists of three parts: a generator set, a data aggregator and a shared discriminator, wherein: the generator set consists of multiple conditional generative adversarial network generators, each task has a conditional generative adversarial network generator, which is used to generate task-related content; the data aggregator is used to summarize the metadata completion relations generated by sampling from multiple task subgraphs; the shared discriminator is a shared component, and by sharing a discriminator, the conditional generative adversarial network discriminator is learned between multiple tasks; Step S1.2.3, construct a meta-learner of the meta-learning framework: construct a meta-learner of the meta-learning framework by accumulating the meta-gradient formed by the objective function of the conditional generative adversarial network and the gradient of each task.
6. The generative adversarial meta-learning knowledge graph completion method for large-scale complex devices according to claim 1 is characterized in that The specific steps of step S1.2.2 are as follows: (1) Based on the differences in graph structures and features of different tasks, the large-scale knowledge graph is decomposed into multiple task sub-graphs. Each task sub-graph is treated as an independent task, and its corresponding conditional generative adversarial network generator is responsible for generating task-related content. (2) Design a conditional generative adversarial network generator for each task subgraph to form a generator set. Each conditional generative adversarial network generator generates entity and relationship vectors related to the task by receiving the conditional input of the corresponding task subgraph. (3) A data aggregator is introduced to aggregate metadata completion relations generated from multiple task subgraphs. The aggregated data is passed to the conditional generative adversarial network discriminator to evaluate the authenticity of the generated content. (4) A shared discriminator component is introduced to achieve learning sharing among multiple tasks, thereby improving the performance and robustness of the conditional generative adversarial network discriminator.
7. The generative adversarial meta-learning knowledge graph completion method for large-scale complex devices according to claim 1 is characterized in that The specific steps of step S3 are as follows: Step S3.1, independently optimizing the base learners in the multi-task generation model; Step S3.2, enabling the meta-learner to achieve migration between multiple related reasoning tasks through global gradient update, so as to quickly adapt to new tasks; Step S3.3: Adjust the multi-task generation model through the multi-task generation model training strategy, decouple the optimization of the multi-task generation model, and adaptively adjust the objective function and the knowledge graph completion strategy.
8. The generative adversarial meta-learning knowledge graph completion method for large-scale complex devices according to claim 1 is characterized in that The specific steps of step S3.1 are as follows: In a specific task, the base learner is trained using limited known information, and the conditional generative adversarial network generator and conditional generative adversarial network discriminator corresponding to the specific task update their loss functions by alternately optimizing, where: when updating the conditional generative adversarial network discriminator, its performance is improved by maximizing the objective function; and when updating the conditional generative adversarial network generator, the performance of the generator is optimized by minimizing the objective function.
9. The generative adversarial meta-learning knowledge graph completion method for large-scale complex devices according to claim 1 is characterized in that The specific steps of step S3.2 are as follows: (1) Using the back-propagation algorithm, the generator loss function is used to calculate the gradient of the parameters of the conditional generative adversarial network generator in each task; (2) Calculate the gradient of the generator loss function with respect to the initial parameters of the meta-learning framework. That is, the calculation of the meta-gradient is based on the gradients of multiple tasks and accumulates them. (3) By utilizing meta-gradients to update the initial parameters of the meta-learning framework, global gradient updates are achieved, enabling the meta-learner to achieve effective transfer learning between multiple related reasoning tasks.
10. The generative adversarial meta-learning knowledge graph completion method for large-scale complex devices according to claim 1 is characterized in that In step S3.3, the multi-task generation model training strategy consists of two parts: local constraints and global gradients.