An independent knowledge graph vector migration method and system based on an evolutionary algorithm

Abstract

This invention belongs to the field of knowledge graph vector transfer technology, and provides an independent knowledge graph vector transfer method and system based on evolutionary algorithms. The method involves: acquiring the source knowledge graph and its trained embedding models; adding all source embedding models to a model pool and evaluating each model using a validation set of the target knowledge graph to select the best-performing model; performing crossover and mutation operations on each model in the model pool and the best-performing model to generate new offspring models; evaluating the offspring models using the validation set to select the current best-performing model; and using the current best-performing model to transfer the source knowledge graph vectors to the target knowledge graph. This invention provides a highly promising research direction for adaptive representation learning in dynamic knowledge graph scenarios.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph vector transfer technology, specifically relating to an independent knowledge graph vector transfer method and system based on an evolutionary algorithm. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Knowledge graphs organize entities and their relationships in the real world in a structured manner, representing facts through triples (subject entity, relation, object entity). Therefore, knowledge graphs provide a powerful representation for the organization and utilization of symbolic knowledge. To address its inherent discreteness, knowledge graph embedding maps entities and relations to a computable vector space. By capturing fundamental representations that reveal latent semantic and structural information, knowledge graphs are promoted for application in downstream tasks such as question-answering systems, recommendation systems, and mitigating the illusion of large language models.

[0004] Cross-knowledge graph embedding transfer aims to effectively transfer shareable knowledge and capabilities learned in a source knowledge graph to a new target knowledge graph. However, in practice, traditional embedding transfer methods focus on static inductive scenarios, where the entities and relations in the target graph are already visible during the training phase of the source knowledge graph. In this case, the transferred content mainly consists of the already learned entity embeddings and relation embeddings. Significant challenges arise when the target knowledge graph introduces unseen entities or relations, such as in inductive scenarios where the target knowledge graph introduces new entities based on relational patterns in the source knowledge graph. For this type of inductive transfer, source knowledge graph entity embeddings cannot be directly applied to new target entities, and the transferable information mainly comes from shared relational patterns. However, existing methods all assume an explicit association between the source and target knowledge graphs, which limits their application in emerging knowledge graph scenarios without any prior association (referred to as independent knowledge graphs). Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes an independent knowledge graph vector transfer method and system based on an evolutionary algorithm.

[0006] According to some embodiments, the present invention adopts the following technical solution: An independent knowledge graph vector transfer method based on an evolutionary algorithm includes the following steps: Obtain the source knowledge graph and its trained embedding model; All source embedding models are added to the model pool, and each model is evaluated using the validation set of the target knowledge graph. The best performing model is selected. Each model in the model pool is combined with the optimal model to perform crossover and mutation operations to generate new offspring models; The offspring model is evaluated using a validation set, and the current best model is selected. The source knowledge graph vectors are then transferred to the target knowledge graph using the current best model.

[0007] As an alternative implementation, all source embedding models are added to the model pool, and each model is evaluated using the validation set of the target knowledge graph. In the process of selecting the best performing model, each embedding model is regarded as an organism, the parameters are corresponding to genes, and the model performance is regarded as a hereditary trait.

[0008] As an alternative implementation, the process of performing a crossover operation on each model in the model pool together with the optimal model includes: generating offspring by combining the genetic information of selected parents, given a pair of parents whose genes, i.e., model parameters, are respectively... and The genetic module uses the crossover operator to generate offspring. Formal representation:

[0009] in, This indicates the proportion of genes inherited from the parent generation. Corresponding to the proportion inherited from the mother generation, extended arithmetic crossover is used, through uniform distribution. The proportion of sampling inheritance, among which Controlling the extent to which offspring parameters exceed the range of parent parameters, when At that time, the method degenerates into standard arithmetic cross.

[0010] As an alternative implementation, the process of performing mutation operations on each model in the model pool together with the optimal model includes: for each offspring individual Through two random rates , Perform the mutation operation, with the mutation rules as follows:

[0011] in, With zero mean and variance random Gaussian perturbation, , The first and second models of the parent and mother models respectively One parameter, The parameter-wise difference between two parent models is defined as a direction vector in the parameter space, which captures the genetic search direction determined by parental diversity. This is the orientation coefficient, used to scale the contribution weight of the parent parameter difference term; This represents the overall mutation probability of offspring individuals. For the first generation of offspring individuals The probability of single-gene mutation in each gene; , These are the threshold values ​​corresponding to the two.

[0012] As a further defined implementation method, , All are predefined fixed values. , Generated through random operations.

[0013] As a further defined implementation method, when and At that time, the first generation of offspring A gene will undergo a mutation operation; otherwise, the gene will not mutate.

[0014] As an alternative implementation method, the process of evaluating the offspring model using a validation set and selecting the current optimal model includes: generating multiple offspring individuals after consecutively performing crossover and mutation operations. Select the individuals with the best performance. :

[0015] in, This indicates that the best-performing offspring are selected from the offspring set based on their evaluation scores. Individual; This is an evaluation function used to measure offspring individuals. Performance on the validation set.

[0016] As an alternative implementation method, during the process of selecting the current optimal model, when the number of models exceeds a predefined threshold... When this happens, models with performance below the set value are removed.

[0017] An independent knowledge graph vector transfer system based on an evolutionary algorithm, comprising: The source knowledge graph data acquisition module is configured to acquire the source knowledge graph and its trained embedding model; The model pool building module is configured to add all source-embedded models to the model pool and evaluate each model using the validation set of the target knowledge graph to select the best performing model. The crossover and mutation module is configured to perform crossover and mutation operations on each model in the model pool together with the best model to generate new offspring models. The optimal model selection module is configured to evaluate the offspring model using a validation set, select the current optimal model, and use the current optimal model to transfer the source knowledge graph vector to the target knowledge graph.

[0018] An electronic device includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, wherein the computer instructions, when executed by the processor, perform the steps in the method described above.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention abstracts each embedding model as an individual composed of genes and models embedding transfer as a genetic evolutionary process in parameter space. Through three core modules—crossover, mutation, and selection—this invention achieves structured parameter recombination from an existing source knowledge graph to a new target knowledge graph in a lightweight, training-free manner. Extensive experiments on multiple target knowledge graphs demonstrate that this invention exhibits superior transfer performance compared to various strong baseline methods. Furthermore, this invention is robust to the selection of source knowledge graphs: regardless of the degree of correlation between the source and target graphs, the framework can effectively utilize the implicit structural patterns contained in the pre-trained parameters through evolutionary recombination. This invention provides a highly promising research direction for adaptive representation learning in dynamic knowledge graph scenarios.

[0020] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0021] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0022] Figure 1 This is an example of an independent knowledge graph vector transfer system architecture based on an evolutionary algorithm. Figure 2 This is a schematic diagram of a target knowledge graph in one embodiment; Figure 3 This is a schematic diagram of a source knowledge graph in one embodiment; Figure 4 This is an experimental result of a method in one embodiment on a baseline migration pattern; Figure 5 This is an experimental result of different source knowledge graphs in one embodiment. Detailed Implementation

[0023] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0025] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0026] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0027] Example 1 In the dynamic evolution of knowledge graphs, the isolation caused by the lack of alignment information between different graphs, and the data sparsity caused by insufficient distribution of triples, have become the core bottlenecks restricting the performance of knowledge embedding learning and downstream reasoning tasks. In practical applications, knowledge graph migration is often required.

[0028] Independent Scenarios within the Same Domain: Taking smart campus construction as an example, although different universities are in the same field, their personnel composition, curriculum system, and organizational structure are highly exclusive, resulting in significant differences in the entities and relationships of their constructed knowledge graphs. For example, different universities may name the same type of course or institution significantly differently: a course called "Computer Application Fundamentals" at University A may be named "University Computer Fundamentals" at University B; the "School of Information Engineering" at University A may correspond to the "Department of Intelligent Science and Engineering" at University B. These entities with similar semantics but completely different identifiers make it impossible to directly reuse knowledge graphs between universities, and it is also difficult to quickly complete the knowledge graphs of newly built universities with the help of external data, further exacerbating the cold start dilemma.

[0029] Cross-domain heterogeneous scenarios: In the cross-domain interaction between industrial equipment operation and maintenance and medical equipment management, the entity semantics (such as "turbine" and "MRI machine") and relationship types of the two are completely different. Turbines in industrial equipment exhibit a correlation pattern of "abnormal speed—increased vibration—decreased output power," while MRI machines in medical equipment correspond to a reasoning link of "magnetic field fluctuation—decreased signal-to-noise ratio—deteriorated imaging quality." Although their domain semantics are completely different, they share a structural pattern of "equipment status—abnormal characteristics—functional impact."

[0030] During migration, only a small amount of target knowledge graph data is needed as a migration guide. For example, a newly built medical equipment management knowledge graph only labels a small number of basic entities such as equipment and faults, along with their relationships. Using this as a migration supervision signal, the structural pattern of "equipment - abnormal characteristics - fault impact" from a mature industrial equipment operation and maintenance graph can be migrated, automatically completing the missing fault associations and transmission paths for medical equipment, without relying on a large amount of target domain data.

[0031] An evolutionary algorithm-based vector transfer method for independent knowledge graphs enables feature sharing between independent knowledge graphs, such as... Figure 1 As shown, this invention uses an analogy with biology: each embedded model is considered an organism, parameters correspond to genes, and model performance is considered a hereditary trait. When faced with a new knowledge graph, the population is initialized using a pre-trained model from the existing knowledge graph. Iterative evolution is achieved through crossover and mutation operations using a genetic algorithm, and dominant individuals are selected based on the performance of the new knowledge graph. Specifically, the crossover module uses extended arithmetic crossover to recombine parent parameters, the mutation module introduces random perturbations to maintain population diversity, and the selection module uses knowledge graph-specific evaluation criteria to guide the evolutionary process. Notably, this invention only requires the target knowledge graph validation set to participate in the computation, and its size is much smaller than the training set, achieving significant computational savings compared to traditional strategies that rely on a complete training set.

[0032] The implementation of this invention relies on the genetic module, namely the core genetic operator, which includes three modules: crossover, mutation, and selection.

[0033] The crossover module implements genetic recombination operations, primarily crossover, which generates offspring by combining the genetic information of selected parents. Given a pair of parents, their genes (i.e., model parameters) are... and The genetic module uses the crossover operator to generate offspring. Formal representation: (1) in, This indicates the proportion of genes inherited from the parent generation, while This corresponds to the proportion inherited from the parent generation. To alleviate the convex hull limitation of standard arithmetic cross, this invention employs extended arithmetic cross, by starting from a uniformly distributed... The proportion of sampling inheritance, among which Control the extent to which offspring parameters exceed the range of parent parameters. When At this point, the method degenerates into standard arithmetic crossover. Larger... This encourages more aggressive exploration but may introduce higher instability. Intuitively, by sampling a random inheritance ratio from a uniform distribution, unbiased random combinations of parental genes can be achieved, while allowing offspring parameters to exceed the range of parents, thus encouraging more diverse structural mixtures.

[0034] The mutation module introduces small and random changes to the individual genome, maintaining population diversity and enhancing the ability to explore the solution space. For each offspring individual... The mutation module uses two random rates , Perform the mutation operation, and the mutation rules are as shown in formula (2): (2) in, With zero mean and variance The random Gaussian perturbation introduces random exploration capability into the algorithm, which helps maintain population diversity. In addition to random noise, this invention further introduces a directed mutation mechanism to optimize and improve the mutation module. Where... , The first and second models of the parent and mother models respectively One parameter, The parameter-wise difference between two parent models is defined as a direction vector in the parameter space, which captures the genetic search direction determined by parental diversity. This is the orientation coefficient, used to scale the contribution weight of the parent parameter difference term.

[0035] This represents the overall mutation probability of offspring individuals. For the first generation of offspring individuals The probability of single-gene mutation in each gene; , The thresholds for both are predefined fixed values ​​in this study. , It can be generated through random operations. When and At that time, the first generation of offspring Each gene will undergo mutation; otherwise, the gene will not mutate. Although the entire genetic evolution process is initialized from the parent model, the extended crossover operator and directed mutation mechanism proposed in this invention can explicitly support offspring individuals to explore parameter space regions beyond the parameter range of the parent model, thereby alleviating the problem of excessive dependence of the model on the parameter boundaries of the source domain.

[0036] The selection module is responsible for screening individuals to participate in reproduction and retention. By prioritizing the retention of individuals with superior performance, it applies genetic pressure, thereby guiding the population to evolve towards the optimal or near-optimal solution. Multiple offspring individuals are generated after the crossover and mutation operations are executed consecutively. The diversity of offspring individuals stems from adaptive genetic variation within the population in response to environmental changes. This characteristic ensures the algorithm's comprehensive exploration of the solution space and is crucial for avoiding getting trapped in local optima. The core objective of the selection module is to design reasonable evaluation metrics to screen out the individuals with the best performance. The calculation formula is shown in (3): (3) in, This indicates that the best-performing offspring are selected from the offspring set based on their evaluation scores. Individual; This is an evaluation function used to measure offspring individuals. Performance on the validation set. It is important to note that the selection process in this invention is performed only on the validation set, not the training set. This significantly reduces computation time compared to methods that fine-tune on a larger training set. Furthermore, performing the selection operation on the validation set effectively mitigates information leakage and reduces the risk of model overfitting: offspring individuals are evaluated and compared on unseen validation data without incurring the computational cost of gradient-based updates. During the prediction phase, the test set will be used to evaluate the performance of the offspring individuals, i.e., to perform… operate.

[0037] Supported by the aforementioned genetic modules, this invention proposes a genetic strategy to transfer embedding representations from existing knowledge graphs to new knowledge graphs. Existing methods typically retrain the model from scratch for new target knowledge graphs, which is not only computationally expensive and time-consuming but also easily limited by the data sparsity inherent in new knowledge graphs. To address these issues, this invention proposes a genetic strategy that, in a lightweight and generalizable manner, transfers the embedding capabilities inherent in the model parameters from a well-trained source knowledge graph to a completely new target knowledge graph.

[0038] Given a set of source knowledge graphs and its trained embedding model The genetic strategy of this invention is designed for each target knowledge graph Learning a high-quality embedding model First, all source embedding models are added to the model pool, and the target knowledge graph is used. Validation set Each model is evaluated, and the model with the best performance is selected. Subsequently, each model in the model pool is compared with the optimal model. Both crossover (Formula 1) and mutation (Formula 2) operations are performed to generate new offspring models. .

[0039] Using the validation set offspring model The model pool is then evaluated to select the current best model. Through the crossover, mutation, and selection operations described above, new offspring models are continuously generated, expanding the model pool. When the number of models exceeds a predefined threshold... At the same time, poorly performing models are removed to maintain the quality of the model pool. Finally, the above genetic steps are executed multiple times to obtain the optimal model suitable for the target knowledge graph, and it is then tested on the target test set. The model is evaluated. After all genetic iterations are completed, the best-performing model and its test results are returned. The model obtained after completing the genetic iterations is the final model after knowledge transfer.

[0040] In summary, the evolutionary process proposed in this invention uses the embeddings learned from the source knowledge graph as the initial genetic seed, and completes iterative evolution through crossover, mutation, and selection operations. Throughout the evolutionary process, only the validation set of the target knowledge graph is used, and its size is much smaller than the training set. Therefore, compared with training strategies that rely on complete target training data, this invention significantly reduces computational overhead.

[0041] This invention proposes a genetically inspired strategy for efficient and transferable embeddings between independent knowledge graphs. Each embedding model is abstracted as an individual composed of genes, and embedding transfer is modeled as a genetic evolutionary process in parameter space. Through three core modules—crossover, mutation, and selection—this invention achieves structured parameter recombination from existing source knowledge graphs to new target knowledge graphs in a lightweight and training-free manner. Extensive experiments on 27 target knowledge graphs demonstrate that this invention exhibits superior transfer performance compared to various strong baseline methods. Furthermore, this invention is robust to the selection of source knowledge graphs: regardless of the degree of correlation between the source and target graphs, the framework can effectively utilize the implicit structural patterns inherent in the pre-trained parameters through evolutionary recombination. This invention provides a highly promising research direction for adaptive representation learning in dynamic knowledge graph scenarios.

[0042] This invention has validated its effectiveness on independent type knowledge graphs. Since there is currently no publicly available dedicated dataset for independent knowledge graphs, an independent dataset was constructed from two dimensions: general domain and specific domain. The specific settings are as follows: 27 target knowledge graphs: 11 domain-specific knowledge graphs and 16 general knowledge graphs were collected as target knowledge graphs, such as... Figure 2 As shown.

[0043] Knowledge graphs for 11 specific domains: Art, Award, Edu, Health, Infra, Loc, Org, People, Science, Sport, and Tax.

[0044] Sixteen general target knowledge graphs: including DBPedia-15K, Wikidata-15K, and YAGO-15K sampled from the OpenEA library, and FB-{25,50,75,100}, WK-{25,50,75,100}, and NL-{0,25,50,75,100} as target knowledge graphs for fully inductive knowledge graphs.

[0045] 23 Source Knowledge Graphs: Adhering to the principle of convenience and accessibility, this invention uses publicly available benchmark knowledge graphs as source knowledge graphs, such as... Figure 3 As shown, the source knowledge graph contains 9 deductive knowledge graphs and 14 inductive knowledge graphs.

[0046] Nine direct-access knowledge graphs: WN18RR, NELL-995, WDsinger, NELL23k, FB15k-237-{10,20,50}, ConceptNet, and CoDEx.

[0047] Fourteen inductive knowledge graphs: FB-V1, FB-V2, FB-V3, FB-V4, NELL-V1, NELL-V2, NELL-V3, NELL-V4, WN-V1, WN-V2, WN-V3, WN-V4, as well as the small-sample version ILPC-S and the large-sample version ILPC-L.

[0048] For each target knowledge graph, all 23 source knowledge graphs are used to guide the genetic algorithm. First, embedding models are pre-trained on the 23 source knowledge graphs, and these pre-trained models are used as the initial seeds for the genetic algorithm. Model selection is guided by the performance of the models on the validation set of the target knowledge graph, and finally, the optimal model is evaluated on the test set of the target knowledge graph.

[0049] It is worth emphasizing that this invention focuses on a training-free transfer learning setting: in this setting, model selection is completed using only validation data without using any target domain training data.

[0050] Evaluation metrics are strongly correlated with specific task types. For the link prediction task of knowledge graphs, the core evaluation metrics include MRR and H@N. To improve the effectiveness of the selection process, five simple evaluation strategies were used to comprehensively consider these two metrics, as shown in formula (4): (4) In the formula , Representing offspring individuals MRR and H@N scores, These are the weighting coefficients. The specific logic for each strategy is as follows: MRR strategy: Measure the overall ranking quality by calculating the inverse mean of the rankings of the correct entity among all candidate entities; H@N strategy: Focus on whether the correct entity appears in the first N prediction results, and measure the top hit rate; Both strategy: MRR and H@N are used as joint core indicators, and the "AND" selection criterion is adopted: only when the candidate offspring individual is better than the current best solution in both MRR and H@N is it determined that its performance is better; Weighted mean strategy: By using a weighted linear combination, MRR and H@N are aggregated into a single scalar target value, which supports a controllable trade-off between the two indicators; Geometric mean strategy: This strategy multiplies and combines MRR and H@N (taking the square root). By penalizing low scores on either indicator, it tends to select individuals with balanced performance on both indicators. Only one evaluation strategy is used per experiment.

[0051] To verify the effectiveness of the present invention, several strong zero-sample migration baselines and representative variants of the present invention were introduced for comparison.

[0052] Baseline methods: This group includes ULTRA and KGICL, which are foundational embedding models applicable to any knowledge graph. These foundational models aim to capture general relational patterns and are pre-trained on large-scale knowledge graph ensembles; the resulting pre-trained models are directly applied to the target knowledge graph for evaluation without any subsequent adaptation.

[0053] Source Domain Baseline Only: This group examines the transferability of knowledge across different knowledge graphs. Embedding models are trained solely on the source knowledge graph and then evaluated directly on the target knowledge graph, without using any target domain data. This group includes 9 transductive models and 14 inductive models, trained on their respective datasets. For example, E(WN18RR) is an embedding model trained on the WN18RR dataset and evaluated directly on each target knowledge graph without further adaptation.

[0054] Mean baseline method: This group uses the method to verify the effectiveness of the proposed evolutionary algorithm by comparing it with the parameter mean baseline. This includes: Uniform mean: uniformly averaging the parameters of all source knowledge graph models; Weighted mean: averaging the parameters according to performance weights based on the validation results on the target knowledge graph.

[0055] GAKG Variants: This group presents the GAKG model proposed in this paper and its representative variants based on different source knowledge graphs: GAKG: the default version, using 9 transductive knowledge graphs and 14 inductive knowledge graphs as source knowledge graphs; GAKG(Trans): using only 9 transductive knowledge graphs as source knowledge graphs; GAKG(Ind): using only 14 inductive knowledge graphs as source knowledge graphs.

[0056] The results of GAKG compared to the baseline migration model are as follows: Figure 4 As shown in the figure, the proposed GAKG framework consistently outperforms baseline methods on the vast majority of datasets, covering both general and domain-specific knowledge graphs. Compared to the current state-of-the-art baseline KGICL, GAKG achieves a stable improvement in average performance across all datasets, with improvements of 2.11, 2.33, 1.98, 1.89, 1.81, and 1.77 in MRR, H@1, H@5, H@10, H@50, and H@100, respectively. Furthermore, paired t-tests confirm that the performance improvement brought by GAKG is statistically significant.

[0057] Unlike basic strategies that rely on large-scale training data and gradient-based optimization, GAKG employs a genetic algorithm paradigm, requiring only a small amount of target domain data to complete model selection. Therefore, the GAKG framework proposed in this invention eliminates the need for training and gradient-based adaptation in the target domain, achieving lightweight embedding transfer while demonstrating significant performance in zero-shot transfer scenarios.

[0058] Compared to experimental results using only the source domain baseline, GAKG yields the following results: Figure 5 As shown in the figure, 9 transductive models and 14 inductive models were trained on their respective datasets and evaluated on the target knowledge graph. For example, E(WN18RR) was trained on the WN18RR training set and achieved an H@1 score of 30.3 on the target knowledge graph DBPedia-15K and 24.7 on the target knowledge graph Wikidata-15K. The performance of GAKG variants using different source knowledge graphs was compared. The average scores of GAKG(Trans), GAKG(Ind), and GAKG(All) models trained using different sets of source knowledge graphs were 34.2, 34.5, and 34.7, respectively. This indicates that GAKG(All), utilizing all available source knowledge graphs, slightly outperforms the GAKG(Trans) and GAKG(Ind) variants that only use a subset of sources.

[0059] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0060] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0061] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0062] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An independent knowledge graph vector transfer method based on evolutionary algorithm, characterized in that, Includes the following steps: Obtain the source knowledge graph and its trained embedding model; All source embedding models are added to the model pool, and each model is evaluated using the validation set of the target knowledge graph. The best performing model is selected. Each model in the model pool is combined with the optimal model to perform crossover and mutation operations to generate new offspring models; The offspring model is evaluated using a validation set, and the current best model is selected. The source knowledge graph vectors are then transferred to the target knowledge graph using the current best model.

2. The independent knowledge graph vector transfer method based on evolutionary algorithm as described in claim 1, characterized in that, In the process of adding all source embedding models to the model pool and evaluating each model using the validation set of the target knowledge graph to select the best performing model, each embedding model is regarded as an organism, with parameters corresponding to genes and model performance as hereditary traits.

3. The independent knowledge graph vector transfer method based on evolutionary algorithm as described in claim 1, characterized in that, The process of performing a crossover operation between each model in the model pool and the optimal model includes: generating offspring by combining the genetic information of selected parents, given a pair of parents whose genes, i.e., model parameters, are as follows: and The genetic module uses the crossover operator to generate offspring. Formal representation: in, This indicates the proportion of genes inherited from the parent generation. Corresponding to the proportion inherited from the mother generation, extended arithmetic crossover is used, through uniform distribution. The proportion of sampling inheritance, among which Controlling the extent to which offspring parameters exceed the range of parent parameters, when At that time, the method degenerates into standard arithmetic cross.

4. The independent knowledge graph vector transfer method based on evolutionary algorithm as described in claim 1, characterized in that, The process of performing mutation operations on each model in the model pool together with the optimal model includes: for each offspring individual Through two random rates , Perform the mutation operation, with the mutation rules as follows: in, With zero mean and variance random Gaussian perturbation, , The first and second models of the parent and mother models respectively One parameter, The parameter-wise difference between two parent models is defined as a direction vector in the parameter space, which captures the genetic search direction determined by parental diversity. This is the orientation coefficient, used to scale the contribution weight of the parent parameter difference term; This represents the overall mutation probability of offspring individuals. For the first generation of offspring individuals The probability of single-gene mutation in each gene; , These are the threshold values ​​corresponding to the two.

5. The independent knowledge graph vector transfer method based on evolutionary algorithm as described in claim 4, characterized in that, , All are predefined fixed values. , Generated through random operations.

6. The independent knowledge graph vector migration method based on an evolutionary algorithm according to claim 5, characterized in that, when and At that time, the first generation of offspring A gene will undergo a mutation operation; otherwise, the gene will not mutate.

7. The independent knowledge graph vector transfer method based on evolutionary algorithm as described in claim 1, characterized in that, The process of evaluating the offspring model using a validation set and selecting the current optimal model includes generating multiple offspring individuals after consecutively performing crossover and mutation operations. Select the individuals with the best performance. : wherein, represents selecting the best performing individuals from the offspring set according to the evaluation score; is an evaluation function for measuring the performance of offspring individuals on the validation set.

8. The independent knowledge graph vector migration method based on evolutionary algorithm according to claim 1, characterized in that, During the selection of the current optimal model, when the number of models exceeds a predefined threshold... When this happens, models with performance below the set value are removed.

9. An independent knowledge graph vector transfer system based on an evolutionary algorithm, characterized in that, include: The source knowledge graph data acquisition module is configured to acquire the source knowledge graph and its trained embedding model; The model pool building module is configured to add all source-embedded models to the model pool and evaluate each model using the validation set of the target knowledge graph to select the best performing model. The crossover and mutation module is configured to perform crossover and mutation operations on each model in the model pool together with the best model to generate new offspring models. The optimal model selection module is configured to evaluate the offspring model using a validation set, select the current optimal model, and use the current optimal model to transfer the source knowledge graph vector to the target knowledge graph.

10. An electronic device, characterized in that, It includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor, which, when executed by the processor, perform the steps of the method according to any one of claims 1-8.

Images