Multi-source data and knowledge graph construction method and system based on learning and adversarial training

Abstract

The invention provides a multi-source data and knowledge graph construction method and system based on learning and adversarial training, and relates to the technical field of enterprise-level technical risk early warning and decision support. According to the method, firstly, a double-path label generation strategy is adopted locally at each client to process data and construct knowledge fragments, so that comprehensiveness and accuracy of knowledge extraction are ensured; then, through an adversarial training mechanism under a federated learning framework, by utilizing the game of an embedded generator and a domain discriminator and the technical implementation of a gradient inversion layer, enabling the model to learn semantic representation with an invariant domain; the method comprises the following steps: optimizing a joint loss function comprising task loss, adversarial loss and semantic alignment loss; and finally, the embedded generator maps the structured knowledge fragments of the clients to a unified semantic space. According to the method, knowledge dispersed in a public domain and a private domain can be effectively integrated under the condition that original data is not leaked, and the knowledge graph which is wide in coverage, consistent in semantics and high in interpretability is constructed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of enterprise-level technical risk early warning and decision support, and in particular to a method and system for constructing multi-source data and a knowledge graph based on learning and adversarial training. BACKGROUND

[0002] With the rapid development of big data and artificial intelligence technology, integrating Internet public data and enterprise internal private data is crucial for scientific research and intelligent decision-making. However, existing data integration and knowledge graph construction technologies face serious challenges. Centralized processing methods require data from various sources to be collected on a central server for unified processing, which seriously violates the requirements of data privacy laws such as the Data Security Law, the Personal Information Protection Law, and GDPR, as sensitive enterprise core data and user privacy data are not allowed to leave the local environment. Although traditional federated learning solves the privacy protection problem of "data not leaving the domain" by training models locally and uploading only model parameters, its optimization goal is usually limited to improving the performance of a single task. When the data sources differ greatly, the model is prone to overfitting to the local data distribution of each client, resulting in learned feature representations that are not universally applicable across domains, lacking domain generalization ability. Existing knowledge graph construction techniques mostly rely on centralized and unified data sources. In cross-public-private-domain scenarios, due to the widespread existence of semantic heterogeneity, i.e., the same concept has different representations in different domains, and strict privacy constraints, it is difficult to directly apply these techniques to construct a unified and integrated knowledge graph.

[0003] In summary, existing technologies cannot simultaneously meet the following three core requirements: strict data privacy protection, i.e., ensuring that data does not leave the local environment; strong generalization ability for multi-source heterogeneous data with large distribution differences, i.e., the model can adapt to different domain data characteristics; and cross-domain semantic alignment and knowledge fusion, i.e., effectively integrating knowledge fragments with the same semantics but different representations scattered in different domains. Specifically, centralized methods violate privacy red lines and cannot meet data security compliance requirements; traditional federated learning has difficulty overcoming domain shift problems, resulting in poor model performance in new domains or cross-domain scenarios; and scattered, semantically inconsistent data islands make it extremely difficult to construct a unified, high-quality knowledge graph, severely restricting the mining and utilization of cross-domain data value. Therefore, there is an urgent need for a technical solution that can effectively address domain differences and semantic heterogeneity while strictly protecting data privacy, enabling multi-source data collaborative fusion and knowledge graph construction. SUMMARY

[0004] The present application aims to provide a method and system for constructing multi-source data and a knowledge graph based on learning and adversarial training to solve the problems existing in the prior art.

[0005] A multi-source data and knowledge graph construction method based on federated learning and adversarial training, comprising the following steps: S1, label generation and knowledge fragment construction: each client uses a double-path label generation strategy to process local data, the double-path includes a keyword-driven path and a path based on a natural language processing model, and generates structured knowledge fragments in the form of triples (entity 1, relationship, entity 2) or (entity, attribute, attribute value); S2, federated learning initialization: the federated learning coordination end initializes the parameters of the embedding generator G, the task classifier C and the domain discriminator D, and distributes the initial parameters to each client, the embedding generator G is used to map the input data x to the latent feature space to generate the embedding vector z = G(x), and the domain discriminator D is connected to the embedding generator G through a gradient reversal layer; S3, local adversarial training: each client uses local data for training, and updates the model parameters by optimizing the joint loss function, the joint loss function includes task loss , adversarial loss and semantic alignment loss , the total loss function of the embedding generator G is , wherein λ and β are hyperparameters; in each training batch, the parameters of the embedding generator G and the task classifier C are fixed to update the domain discriminator D, and then the parameters of the domain discriminator D are fixed to update the embedding generator G and the task classifier C; S4, parameter aggregation: after each client completes local training, the parameter update amount of the embedding generator G and the task classifier C is calculated and uploaded, the federated learning coordination end uses the federated averaging algorithm to weight and aggregate the collected parameter updates, the aggregation formula is , wherein is the data volume of the i-th client, is the total data volume, and the aggregated global model parameters are distributed to each client; S5, knowledge graph collaborative construction: repeat steps S3 to S4 until the model converges, use the trained embedding generator G to generate domain-invariant embedding vectors for the knowledge fragments of each client, align and cluster similar entities in the global embedding space by calculating the similarity of the embedding vectors, merge similar entities into unified entities, and integrate the relationship information of each client to construct a cross-domain knowledge graph; S6, bloodline tracking: record the data source, processing path, model version and fusion process for each entity and relationship in the knowledge graph, the data source includes domain identifier and client identifier, the processing path includes label generation strategy and generation time, and the fusion process includes aligned entities and alignment algorithm.

[0006] By the technical scheme, the cross-domain knowledge graph collaborative construction under the premise of strictly protecting data privacy is realized through a systematic process of six steps. The method first uses a double-path label generation strategy to process data and construct knowledge fragments locally on each client, ensuring the comprehensiveness and accuracy of knowledge extraction; then, through the adversarial training mechanism under the federated learning framework, the game between the embedding generator and the domain discriminator and the gradient reversal layer technology are used to make the model learn the domain-invariant semantic representation, effectively overcoming the domain shift and semantic heterogeneity problems; by optimizing the joint loss function containing task loss, adversarial loss and semantic alignment loss, the cross-domain semantic fusion is realized while ensuring the performance of downstream tasks; through the federated averaging algorithm, the parameter updates of each client are weighted and aggregated to form a global model and iteratively optimized; finally, the trained domain-invariant embedding generator is used to generate embedding vectors for knowledge fragments, and entity alignment and clustering are performed in the global semantic space to construct a unified cross-domain knowledge graph; and the complete data source and processing path are recorded through the bloodline tracking module. This method can effectively integrate knowledge scattered in public and private domains without revealing the original data, construct a knowledge graph with wide coverage, consistent semantics and strong explainability, and significantly improve the collaborative utilization value of cross-domain data, providing high-quality knowledge services for research analysis, intelligent decision-making, technical intelligence, etc. and meeting the increasingly stringent data security and privacy protection regulations.

[0007] Specifically, the technical effects of the method are reflected in the following aspects: first, the federated learning paradigm is used to realize privacy protection without data leaving the domain, and each client only uploads model parameter updates rather than raw data, fundamentally eliminating the risk of data leakage; second, the domain invariance is realized through the dual mechanisms of adversarial training and semantic alignment, making the domain discriminator unable to distinguish the feature source, while cross-domain semantic similar samples are closely clustered in the embedding space, effectively solving the domain generalization problem of traditional federated learning; third, through double-path label generation and joint loss function optimization, the quality of knowledge extraction and the stability of model training are ensured; fourth, through entity alignment and clustering algorithms, automatic fusion of cross-domain knowledge is realized, significantly reducing the cost of manual intervention; fifth, through the bloodline tracking mechanism, the whole life cycle management of the knowledge graph is realized, enhancing the explainability, credibility and auditability of the system. In summary, this method provides an innovative, feasible and efficient technical solution to the three core problems of privacy protection, semantic fusion and knowledge construction in cross-domain data collaboration.

[0008] Further, the task loss includes a label prediction loss and a knowledge graph construction loss , wherein: the label prediction loss , where N is the number of samples in the batch, K is the total number of label categories, is the true label, is the predicted probability; knowledge graph construction loss , where M is the number of triples, and are the subject entity and object entity, respectively, is the relationship; task loss , where γ is the balance coefficient.

[0009] Further, the adversarial loss L_adv in step S3 is defined as: , where and are the data distributions of the public domain and the private domain, respectively; Adversarial training is achieved through a gradient reversal layer. During forward propagation, the gradient reversal layer acts as an identity mapping. During backward propagation, the gradient from the domain discriminator D is multiplied by -1 and then passed back to the embedding generator G.

[0010] Further, the semantic alignment loss L_sem in step S3 is defined as: The InfoNCE loss of contrastive learning is used for the cross-domain positive sample pair , and the semantic alignment loss is defined as: , where P is the number of positive sample pairs, and are the embedding representations, sim(·,·) is the similarity function, and τ is the temperature coefficient.

[0011] Further, the specific implementation of the dual-path label generation strategy in step S1 is as follows: The keyword-driven path extracts keywords and key phrases from the text as entity labels using the TF-IDF algorithm; The path based on the natural language processing model performs named entity recognition and relationship extraction on the text using a pre-trained language model to generate structured knowledge fragments.

[0012] Further, in step S4, the client adds differential privacy noise to the parameter update before uploading the parameter update. The parameter update after adding noise is , where S is the sensitivity of the parameter update, ε is the privacy budget, and δ is the failure probability of privacy loss.

[0013] Further, the specific steps of the entity alignment in step S5 are: calculating the cosine similarity between any two entity embedding vectors, identifying two entities as synonymous entities when the similarity is greater than a preset threshold, using a clustering algorithm to group entities with high similarity into a group, selecting a standard entity name as a representative, and other names as aliases, and recording the original source information of each entity.

[0014] A multi-source data and knowledge graph construction system based on federated learning and adversarial training, comprising: A federated learning coordination end for initializing global model parameters, distributing model parameters to each client, collecting parameter updates uploaded by each client, using a federated averaging algorithm for parameter aggregation, and distributing the aggregated model parameters to each client; A plurality of clients, each client comprising: An embedding generator G for receiving local data and outputting high-dimensional semantic embedding vectors, implementing a mapping function z=G(x); A domain discriminator D, which is a binary classification neural network, for receiving vectors z generated by the embedding generator G and judging their source domain, the domain discriminator D being connected to the embedding generator G through a gradient reversal layer; A task classifier C, which is a multi-label classification neural network, for receiving vectors z generated by the embedding generator G and outputting prediction results of multiple labels; A label generation module for generating structured knowledge fragments using a double-path label generation strategy, the double path including a keyword-driven path and a path based on a natural language processing model; A local training module for training the embedding generator G, the domain discriminator D, and the task classifier C using local data, and updating model parameters by optimizing a joint loss function including task loss, adversarial loss, and semantic alignment loss; A knowledge graph construction module for generating embedding vectors for knowledge fragments of each client using the trained embedding generator G, performing entity alignment and clustering in a global embedding space, and constructing a cross-domain knowledge graph; A bloodline tracking module for recording the data source, processing path, model version, and fusion process of each entity and relationship in the knowledge graph.

[0015] Further, the embedding generator G adopts a pre-trained language model based on a Transformer architecture, which is fine-tuned during the federated learning process to achieve domain invariance; the domain discriminator D is composed of several fully connected layers, and finally outputs domain classification probability through a sigmoid activation function; the task classifier C is composed of several fully connected layers, and finally outputs independent probability predictions for each label through a sigmoid activation function.

[0016] Further, the information recorded by the bloodline tracking module includes: source domain identifier of the entity, client identifier, label generation strategy, model version, generation time, other entities participating in alignment, alignment algorithm and alignment time; source client of the relationship, extraction method and confidence score.

[0017] Advantages of the present application: (1) The present application realizes strict data privacy protection through the federated learning framework. In the whole training and knowledge graph construction process, the original data of each client is always kept in the local environment and does not need to be uploaded to the central server or other nodes, and only the model parameter update information is exchanged, which fundamentally meets the "data does not leave the domain" regulatory requirements. Compared with the centralized processing method which needs to collect original data, the present application avoids the risk of data leakage and complies with the data privacy protection regulations such as the Data Security Law, the Personal Information Protection Law and GDPR. In addition, the present application also supports adding differential privacy noise before parameter uploading, which further enhances the privacy protection strength, and even if the attacker obtains the uploaded model parameters, it is difficult to infer the original data information, providing reliable technical support for enterprises and institutions to cooperate across domains under the compliance premise.

[0018] (2) The present application effectively solves the problems of domain generalization and semantic heterogeneity by introducing the adversarial training mechanism and semantic alignment loss. The adversarial game between the embedding generator and the domain discriminator forces the model to learn domain-invariant feature representation, so that the data from different domains cannot be distinguished in the feature space, thus eliminating the domain-specific information in the features and retaining the essential semantic content. This domain invariance significantly improves the generalization ability of the model in unknown domains and cross-domain scenarios, avoiding the problem of model overfitting to the local domain data distribution in traditional federated learning. At the same time, the semantic alignment loss explicitly reduces the distance between cross-domain semantic similar samples through contrastive learning, directly solving the semantic heterogeneity problem of inconsistent expression of the same concept in different domains. Experimental results show that the accuracy of the domain discriminator is reduced to about 50%, close to the random level, proving the effective implementation of domain invariance; at the same time, the similarity threshold of cross-domain entity alignment can reach more than 0.9, which is significantly higher than the traditional method without using adversarial training and semantic alignment, fully verifying the superiority of the present application in handling multi-source heterogeneous data.

[0019] (3) The application significantly enhances the explainability, credibility and auditability of the knowledge graph through the built-in data bloodline tracking mechanism. The bloodline tracking module records the complete life cycle information of each entity, relationship and label in the knowledge graph, including the domain identifier and client identifier of the data source, the processing steps and model version used for data processing, the algorithm for entity alignment and other entities involved in fusion, and detailed metadata such as the generation time and application scenario of the final knowledge. These bloodline information enables users to trace the complete path from the original data to the final output of any knowledge conclusion, understand the whole process of "where from, how to process, why to merge or label", and realize the transparency and explainability of knowledge. This is particularly important for application scenarios that require strict supervision and review, such as financial risk control, medical diagnosis and other fields, and decision-makers can verify the reliability of the knowledge source and the rationality of the generation process through bloodline tracking. In addition, bloodline tracking also supports version management and quality control of the knowledge graph, when an error is found in a certain knowledge, the source can be quickly located for correction, and the affected downstream knowledge can be tracked to ensure the continuous high quality of the knowledge graph. Compared with the current knowledge graph construction technology which lacks systematic bloodline management, the application provides more perfect and reliable knowledge services. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only illustrate the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0021] Figure 1 The system architecture diagram of the application: shows the overall architecture of the system of the application, including the federated learning coordination end, multiple clients, core components (embedding generator G, domain discriminator D, task classifier C) in each client and bloodline tracking module. The data flow direction of parameter uploading / delivering is clearly marked in the figure; Figure 2 The federated adversarial training flowchart of the application: depicts the calculation process of a single client in a round of training, including the forward propagation of input data through embedding generator G, task classifier C and domain discriminator D, the calculation of three loss functions (L1, L2, L3), and the backward propagation and parameter update path through the gradient reversal layer; 、 、 ; Figure 3The domain-invariant semantic space fusion schematic diagram of the application: taking a two-dimensional vector space as an example, the state of the pre-training public domain and private domain entities gathering in the space and separating from each other, and the effect of the semantic similar cross-domain entities gathering into a group in the domain-invariant semantic space after training to realize fusion are shown. DETAILED DESCRIPTION

[0022] The technical solutions of the application will be described in detail below with reference to the drawings. It should be understood that the embodiments described below are only used to illustrate and explain the application, and are not used to limit the application. The application proposes a multi-source data and knowledge graph construction method and system based on federated learning and adversarial training. The method realizes semantic fusion of cross-domain heterogeneous data and collaborative knowledge construction by learning domain-invariant semantic representation under the premise of strictly protecting data privacy.

[0023] As shown in the accompanying Figure 1 The system architecture of the application includes a federated learning coordination end, a plurality of clients and a bloodline tracking module. Each client includes three core components: an embedding generator G, a domain discriminator D and a task classifier C. The federated learning coordination end is responsible for initializing the global model, distributing the model to each client, aggregating the model parameter updates uploaded by each client, and issuing the aggregated new model to each client. In the whole process, the federated learning coordination end does not touch any original data, only circulates model parameters, thereby protecting data privacy. The data held by each client can be public domain data or private domain data. Public domain data, for example, is public information from the Internet, and private domain data, for example, is sensitive data or user privacy data within an enterprise. These data are processed and trained locally in each client without leaving the local environment.

[0024] The technical feature of the embedding generator G is to receive local data as input and output a high-dimensional semantic embedding vector. Specifically, the embedding generator G implements a mapping function to map the input data x to the latent feature space, i.e. z=G(x). The core role of the embedding generator G is to learn a representation that can strip the data source information, so that the output vector does not contain domain features as much as possible, while retaining rich task semantic information. In actual implementation, the embedding generator G can use a pre-trained language model based on the Transformer architecture, such as BERT or its variants, to obtain initial semantic representation capability through pre-training on a large-scale corpus, and then fine-tune in the federated learning process to adapt to specific task requirements and achieve domain invariance. The parameterization goal of the embedding generator G is to make the generated feature representation z as confusing as possible to the domain discriminator D, i.e. to make the domain discriminator unable to determine whether the feature representation z comes from the public domain or the private domain.

[0025] The domain discriminator D is a binary classification neural network model, which is characterized by receiving the vector z generated by the embedding generator G as input, and judging whether the vector comes from the public domain or the private domain. The domain discriminator D outputs a scalar probability D(z) to estimate the probability that z comes from the public domain. The goal of the domain discriminator D is to accurately distinguish the source domain of the features, which is used to guide the adversarial training. In actual implementation, the domain discriminator D is usually composed of several fully connected layers, and finally outputs a probability value between 0 and 1 through the sigmoid activation function. The domain discriminator D and the embedding generator G are connected through a gradient reversal layer, forming an adversarial relationship. The gradient reversal layer is a special neural network layer that acts as an identity mapping during forward propagation and multiplies the gradient by a negative constant during backward propagation, thereby achieving the effect of adversarial training. In this way, when optimizing the domain discriminator D, the gradient propagates normally to improve the discriminant ability of D; when optimizing the embedding generator G, the gradient is reversed, so that the optimization direction of G is to reduce the discriminant ability of D, that is, to generate feature representations that D cannot distinguish the source.

[0026] The task classifier C is a multi-label classification neural network model, which is characterized by receiving the vector z generated by the embedding generator G as input, and outputting the prediction results of multiple labels. The task classifier C provides task-oriented supervision signals for the embedding generator G to prevent G from losing useful semantic information in order to deceive the domain discriminator D. In actual implementation, the task classifier C is usually composed of several fully connected layers, and finally outputs independent probability predictions for each label through the sigmoid activation function, supporting multi-label classification tasks. The training goal of the task classifier C is to minimize the difference between the predicted labels and the true labels, so as to ensure that the feature representations generated by the embedding generator G are effective for downstream tasks.

[0027] The bloodline tracking module is a record and management system, which is characterized by recording the data source, processing path, model version and fusion process of each entity, relationship or label in the knowledge graph. Specifically, the bloodline tracking module retains complete data bloodline information for each triple or label, including the domain identifier of the source data, the processing steps of data processing, the aggregation version of model parameters, the rules and paths of entity merging, and the consumption scenarios of the final knowledge. These bloodline metadata make the source and generation path of each conclusion in the knowledge graph queryable and auditable, thereby enhancing the explainability and credibility of the system, and meeting the needs of regulatory compliance and decision support.

[0028] As shown in FIG. 1, the system comprises a data preprocessing module 100, an embedding generator 200, a domain discriminator 300, a task classifier 400, a bloodline tracking module 500, and a knowledge graph 600. Figure 2As shown, the training process of the method of the application includes the following steps. First is the label generation stage, which uses a double-path label generation strategy locally on the client side to generate high-quality labels and knowledge fragments for the original data. The first path is a keyword-driven path that extracts keywords and key phrases from the text through rules or statistical methods such as TF-IDF, which can be directly used as entity labels or used to construct simple triples. The second path is a model-driven path based on natural language processing models, which uses pre-trained language models such as BERT to perform deep semantic analysis on the text, identify entities, relationships, and attributes, and generate structured knowledge fragments. The two paths can complement each other, with the keyword-driven path being fast and controllable, suitable for extracting high-frequency and prominent concepts; the model-driven path can capture deep semantics and implicit relationships, suitable for discovering complex and fine-grained knowledge. After fusion and deduplication, the labels and knowledge fragments generated by the two paths form the local knowledge base of the client.

[0029] Next is the knowledge fragment generation stage, based on the results of the double-path label generation, the client further converts the local data into structured knowledge fragments. Knowledge fragments are usually represented in the form of triples, i.e. (entity1, relationship, entity2) or (entity, attribute, attribute value). For example, for a technical paper, multiple triples such as (paper title, author, author name), (author name, affiliation, institution name), (paper title, research field, field name) can be extracted. These knowledge fragments retain the semantic information of the data, but remove the detailed content of the original text, thereby protecting the data privacy to some extent. At the same time, metadata is attached to each knowledge fragment, including source domain identifier, generation time, generation rule or model version, etc., preparing for subsequent bloodline tracking.

[0030] Then enter the local training stage, which is the most core step in the federated learning framework. In each round of federated learning training, the federated learning coordination end first distributes the current global model parameters to all participating clients. Each client receives the global model and performs several rounds of local training using local data. The goal of local training is to optimize a joint loss function, which includes three parts: task loss, adversarial loss, and semantic alignment loss.

[0031] The task loss L_task is used to ensure that the model can effectively complete the downstream task, which includes label prediction loss and knowledge graph construction loss. The label prediction loss L_label is defined as:

[0032] where N is the number of samples in the batch, K is the total number of label categories, is the true value of the kth label of the ith sample, which is 0 or 1, is the predicted probability of the k-th label of the i-th sample by the task classifier C. The loss function adopts the form of binary cross-entropy, which is suitable for multi-label classification tasks, and enables the task classifier to make binary classification judgments on each label independently.

[0033] Knowledge graph construction loss is used to guide the model to generate high-quality knowledge fragments, which are specifically defined as:

[0034] where M is the number of triples in the batch, and are the subject entity and object entity of the j-th triple, is the relationship between them, is the probability that the model predicts the triple to be correct. The loss function guides the model to learn effective representations of entities and relationships by maximizing the likelihood of correct triples. Task loss L_task is the weighted sum of label prediction loss and knowledge graph construction loss, that is:

[0035] where γ is the balance coefficient, used to adjust the relative importance of the two loss terms, and usually takes a value range of 0.1 to 1.0.

[0036] Adversarial loss L_adv is used to realize domain invariance, which makes the generated feature representations indistinguishable between different domains through the adversarial game between embedding generator G and domain discriminator D. Adversarial loss is defined as:

[0037] where and represent the data distribution of the public domain and the private domain, and are data samples sampled from the public domain and the private domain, respectively, and D(G(x)) represents the discrimination result of domain discriminator D on the feature representation generated by embedding generator G. The form of this loss function is similar to the minimax game in the generative adversarial network, and the goal of domain discriminator D is to maximize the value function V(D, G), that is, to improve its ability to distinguish the source domain of the features; the goal of embedding generator G is to minimize the value function V(D, G), that is, to generate features that domain discriminator cannot distinguish the source of.

[0038] The implementation of adversarial training is completed through the gradient reversal layer. In the training process, first fix the parameters of embedding generator G, and update the parameters of domain discriminator D through the standard gradient ascent method, and the optimization goal is:

[0039] where are parameters of the domain discriminator D, is the learning rate, is the gradient of the value function V(D, G) with respect to the parameters of D. Then the parameters of the domain discriminator D are fixed, and the parameters of the embedding generator G are updated by the gradient reversal layer, and the optimization objective is:

[0040] where are parameters of the embedding generator G, is the learning rate. Note that here is gradient descent rather than gradient ascent, because the goal of G is to minimize V(D, G). In actual implementation, by passing the output of the embedding generator G through the gradient reversal layer when forward propagating, and multiplying the gradient from D by -1 before passing it back to G when backward propagating, the above optimization process is automatically implemented.

[0041] Semantic alignment loss is used to explicitly pull the distance between cross-domain semantically similar samples, implemented using the InfoNCE loss of contrastive learning. First, cross-domain positive sample pairs need to be constructed, i.e. find semantically similar samples in the public domain and the private domain. This can be achieved in the following ways: use pre-trained models or rule-based methods to identify cross-domain synonymous entities or related concepts, for example, "artificial intelligence" in the public domain and "artificial intelligence" in the private domain are different expressions of the same concept; or use manual annotation or semi-supervised learning methods to mark a small number of cross-domain correspondence relationships as anchor points. For a cross-domain positive sample pair , the semantic alignment loss is defined as:

[0042] where P is the number of positive sample pairs, P is the set of positive sample pairs are the embedding representations of the public domain sample and the private domain sample respectively, sim(·, ·) is the similarity function, usually cosine similarity, τ is the temperature coefficient used to adjust the smoothness of the distribution, usually taken as 0.05 to 0.5, and the summation in the denominator is performed on all negative samples in the batch. This loss function adopts a symmetric form, considering both the public domain to the private domain and the private domain to the public domain contrastive learning, ensuring bidirectional semantic alignment effect.

[0043] Combining the above three loss terms, the total loss function of the embedding generator G is obtained:

[0044] where λ and β are hyperparameters to balance the relative importance of the three loss terms. λ controls the strength of adversarial training, typically ranging from 0.1 to 1.0, larger λ makes domain invariance stronger but may lose part of task performance. β controls the strength of semantic alignment, typically ranging from 0.1 to 1.0, larger β makes cross-domain sample alignment tighter but needs high-quality positive samples. Note that there is a negative sign in front of the adversarial loss, because the goal of embedding generator G is to minimize the adversarial loss, i.e., to make the discriminative ability of domain discriminator D decline.

[0045] In each training batch, the client updates model parameters in the following order. First, fix the parameters of embedding generator G and task classifier C, update the parameters of domain discriminator D using adversarial loss alone:

[0046] Then fix the parameters of domain discriminator D, update the parameters of embedding generator G and task classifier C using the total loss function:

[0047]

[0048] where is the parameter of task classifier C, is the learning rate. This alternating update approach ensures the stability of adversarial training, avoiding oscillation between generator and discriminator.

[0049] After the client completes local training, it calculates the update amount of model parameters, i.e., the difference between the current parameters and the initial parameters. To further protect privacy, differential privacy noise such as Gaussian noise or Laplace noise can be added to the parameter update amount, and the scale of the noise is controlled by the privacy budget parameter. Then the client uploads the parameter update amount of embedding generator G and task classifier C to the federated learning coordination end, note that the parameters of domain discriminator D are not uploaded because domain discriminator is only used for adversarial training locally and does not participate in global aggregation.

[0050] Enter the parameter aggregation phase, after the federated learning coordination end collects all the parameter updates uploaded by the clients, it uses the federated averaging algorithm or its variants for aggregation. The basic form of the federated averaging algorithm is:

[0051] where is the global model parameter of the t+1 round, n is the number of clients participating in this round of training, is the local model parameter of the i-th client after the t+1 round of training, is the local data volume of the i-th client, is the total data amount of all participating clients. This aggregation method performs a weighted average according to the data amount of each client, and the client with a larger data amount contributes more to the global model. In addition to the federated averaging algorithm, the FedProx algorithm can be used to add a regularization term to constrain the local model from deviating too far from the global model during local training, or the FedAdam algorithm can be used to aggregate with an adaptive learning rate. After aggregation is completed, the federated learning coordinator end distributes the new global model parameters to each client, and the next round of training begins. This process is repeated until the model converges or the preset number of training rounds is reached.

[0052] After multiple rounds of federated adversarial training, the embedding generator G learns a domain-invariant semantic representation, that is, as long as the semantics of the input data from the public domain or the private domain are similar, the generated embedding vectors will be close to each other in the vector space. Using this domain-invariant representation, global semantic space and knowledge graph can be constructed collaboratively. The specific steps are as follows. First, apply the trained embedding generator G to all client knowledge fragments to generate corresponding embedding vectors for each entity and relationship. Since the embedding generator G has domain invariance, synonymous entities from different domains will obtain similar embedding vectors. Then, perform semantic neighbor search or clustering analysis in the global embedding space to identify entities with similar embedding vectors. For example, using K-Means clustering or DBSCAN density clustering algorithm, entities with a distance less than a threshold in the embedding space are clustered into a class, and each class represents a unified semantic concept. For multiple entities in the same class, perform entity alignment and merging operations, and select a standard entity name as the representative, and other names as aliases or synonyms of the entity. In the merging process, the original source information of each entity needs to be preserved, i.e., which domain and which client they come from, for subsequent bloodline tracking.

[0053] After entity alignment, further integrate the relationship information uploaded by each client to construct a cross-domain consistent entity relationship network. For the same entity pair, if there are multiple relationships from different domains, relationship fusion or conflict resolution needs to be performed. A rule-based method can be used, such as preferentially selecting a relationship with high confidence, or using a voting mechanism to select the relationship recognized by the majority of clients; or a model-based method can be used, using knowledge graph embedding techniques such as TransE or RotatE to score the rationality of the relationship, and selecting the relationship with a high score. The final knowledge graph is a cross-domain integrated, semantically unified entity relationship network, with clear semantic definitions and rich structural information for entities and relationships.

[0054] Throughout the construction of the knowledge graph, the bloodline tracking module continuously records key metadata information. For each entity, record its data source including source domain identifier, client identifier and reference of original data, record its generation path including the used label generation strategy, model version and generation time, record its fusion process including other entities involved in alignment, alignment algorithm and alignment time. For each relationship, record its source information, extraction method and confidence. For each label or attribute, record its generation rule and verification status. These bloodline metadata are stored in a special bloodline database associated with the content of the knowledge graph. By querying the bloodline database, the complete life cycle of any element in the knowledge graph can be traced back, from the original data to the whole process of the final knowledge, realizing the transparency and traceability of "where from, how to process, why to merge or label". This is of great significance for regulatory audit, quality control and user trust.

[0055] The constructed knowledge graph can support a variety of downstream applications. In the multi-label profiling application, the rich semantic information of the knowledge graph is used to generate a comprehensive set of labels for users or entities, which is used for precise recommendation or personalized service. In the cross-domain retrieval application, the domain-invariant semantic representation is used to realize unified retrieval across public and private domains, and users can use one query to retrieve both public information and authorized private information. In the technology evolution monitoring and early warning application, by analyzing the trends of entities and relationships in the knowledge graph over time, emerging technologies, hot research directions and potential disruptive innovations are identified, providing support for scientific research decision-making and strategic planning. In the expert and institution recommendation application, based on the person entities, institution entities and their relationship networks in the knowledge graph, relevant experts, scholars or research institutions are recommended to users to promote cooperation and exchange.

[0056] The method of the present application also supports dynamic updating and continuous learning. With the generation of new data or changes in user demand, the client can perform a new round of local training, and the federal learning coordination end can trigger model aggregation and distribution regularly or on demand, thereby realizing the continuous evolution of the knowledge graph. In the incremental learning mode, part of the trained parameters can be fixed and only the newly added part can be updated, improving the training efficiency. At the same time, through the version control mechanism, the models and knowledge graphs of different periods are managed, supporting historical tracing and version comparison, meeting the needs of long-term maintenance and audit.

[0057] Embodiment one The technical solutions of the present application are further illustrated below through a specific embodiment. Assume that a scientific research institution wants to build a cross-domain knowledge graph covering public academic literature and internal research reports for technology intelligence analysis and scientific research project management. The challenge faced by the institution is that public academic literature can be freely accessed from the Internet, but internal research reports involve unpublished scientific research achievements and sensitive information and cannot be uploaded to an external server for centralized processing. By using the method of the present application, semantic fusion and knowledge graph construction of public literature and internal reports can be realized on the premise of protecting the privacy of internal data.

[0058] In the system architecture building stage, a federated learning coordination end server is deployed for managing the global model and coordinating the training process. The server can be deployed on the internal network of the scientific research institution or a trusted cloud platform, but does not store any original data. Two types of clients are set up, the public domain client is deployed on a server with internet access, responsible for collecting and processing public academic literature, such as crawling paper abstracts and metadata from academic databases such as arXiv and PubMed; the private domain client is deployed on the internal server of the scientific research institution, has access to the internal research report database, but is physically isolated from the external network or strictly protected by a firewall to ensure that internal data is not leaked. Each client is installed with three components, an embedding generator G, a domain discriminator D and a task classifier C, as well as related software modules for local training and data processing.

[0059] In the data preparation stage, the public domain client collects 100,000 paper abstracts in the field of artificial intelligence from academic databases, including title, author, abstract text, keywords and other information. The private domain client extracts 5,000 research reports from the internal database, including project name, researcher, report abstract, technical keyword and other information. Note that the data volume of the private domain is relatively small, which is common in actual scenarios, because the accumulation speed of enterprise internal data is usually much lower than that of public data.

[0060] During the tag generation phase, each client employs a dual-path strategy to generate tags and knowledge fragments for the data. In the public domain client, the keyword-driven path uses the TF-IDF algorithm to extract keywords from the paper abstract. For example, for a paper on deep learning, keywords such as "deeplearning," "neural network," and "convolutional" are extracted as technical tags. The model-driven path uses the BERT model to perform named entity recognition and relation extraction on the abstract, identifying entities such as author name, institution name, research method, and application field, as well as relations such as "author-institution" and "method-applied-field," generating triples such as ("John Doe", "works_at", "MIT") and ("deeplearning", "applied_in", "computervision"). In the private domain client, a similar method is used to process internal research reports. However, since the format and terminology of internal reports may differ from public papers—for example, the same concept may be expressed differently—appropriate adjustments or customized rules are needed during keyword extraction and entity recognition. For example, internal reports may use "deep learning" instead of "deeplearning", or "neural network" instead of "neuralnetwork". These synonyms need to be identified and aligned in subsequent semantic alignment.

[0061] During model initialization, the federated learning coordinator initializes the parameters of the embedding generator G, using a pre-trained BERT-based model as the initial weights. This model has been pre-trained on a large-scale general corpus and possesses good semantic representation capabilities. The task classifier C is initialized as a three-layer fully connected network with an input dimension of 768 (corresponding to the hidden layer dimension of BERT) and an output dimension of 500 (corresponding to the predefined number of label categories), using random initialization. The domain discriminator D is initialized as a two-layer fully connected network with an input dimension of 768 and an output dimension of 1 representing the probability of domain classification, also using random initialization. The federated learning coordinator then distributes these initial parameters to all clients.

[0062] During the local training phase, each client receives the initial model and begins local training. The batch size for local training is set to 32, and the learning rate is... =2×10^(-5) is used for embedding generators. =1×10^(-3) is used for task classifiers. = 1 x 10^(-3) for domain discriminator. The hyperparameters of the loss function are set as γ = 0.5 to balance the label prediction and knowledge graph construction, λ = 0.5 to control the adversarial training intensity, and β = 0.3 to control the semantic alignment intensity. In each training batch, the client randomly samples 32 samples from the local data, generates embedding vectors through the embedding generator G, and then inputs them into the task classifier C and the domain discriminator D, respectively. When calculating the task loss , the binary cross-entropy loss is calculated by comparing the real label and the predicted label of the sample. If the sample has corresponding triple information, the knowledge graph construction loss is also calculated. When calculating the adversarial loss , the domain discriminator D judges the source domain of the embedding vector and compares it with the real domain label. The domain label of the sample of the public domain client is 0, and the domain label of the sample of the private domain client is 1. When calculating the semantic alignment loss , cross-domain positive sample pairs need to be constructed. In this embodiment, the positive sample pairs are constructed in the following way: a synonym dictionary is defined in advance, which contains the corresponding relationship of terms in the public domain and the private domain, such as ("deeplearning", "deep learning"), ("neural network", "neural network"), etc.; during training, if there are samples containing these synonyms in the batch, they are paired as positive sample pairs. For the sample containing "deeplearning" in the public domain client and the sample containing "deep learning" in the hypothetical private domain sample, the InfoNCE contrast loss of their embedding vectors is calculated. Since each client can only access local data, certain skills are needed to construct cross-domain positive sample pairs, and methods such as a small amount of manual annotation, rule matching, or pre-trained model generation can be used.

[0063] During back propagation and parameter updating, first, fix the parameters of G and C, update the parameters of the domain discriminator D using the gradient of the adversarial loss, so that D improves the discrimination ability of the domain. Then fix the parameters of D, update the parameters of the embedding generator G using the gradient of the total loss . Due to the effect of the gradient reversal layer, the gradient of the adversarial loss term on G is in the opposite direction, so that G learns to generate features that cannot be distinguished by D. At the same time, the parameters of the task classifier C are updated using the gradient of the task loss, to ensure that C can accurately predict the label. Each client performs local training for 5 epochs, i.e., the local data set is traversed 5 times. During training, the client monitors the changes of the loss function, and when the loss tends to be stable or reaches the preset number of epochs, the local training is stopped.

[0064] In the parameter uploading stage, the client calculates the update amount of the model parameters, i.e. and , where and are the parameters after local training, and are initial parameters issued from the coordination end before training. In order to further protect privacy, Gaussian noise is added to the parameter update amount, and the noise scale is controlled by the privacy budget ε, which is set to 1.0 in this embodiment. The parameter update amount after adding noise is:

[0065] wherein S is the sensitivity of parameter update, and δ = 10^(-5) is the failure probability of privacy loss. The client uploads the parameter update amount after adding noise to the federated learning coordination end, and uploads the sample quantity N i of the local data, but does not upload the parameters of the domain discriminator D.

[0066] In the parameter aggregation stage, the federated learning coordination end collects all the parameter updates uploaded by the clients, and uses the federated averaging algorithm for aggregation. Since the data quantity of the public domain client is 100,000, and the data quantity of the private domain client is 5,000, the total data quantity is 105,000. The weight of the public domain client is 100,000 / 105,000≈0.952, and the weight of the private domain client is 5,000 / 105,000≈0.048. The global model parameters after aggregation are This weighted aggregation method ensures that the client with larger data quantity has greater influence on the global model, but also retains the contribution of a small number of clients, avoiding excessive bias of the global model to one party. After aggregation, the federated learning coordination end updates the global model parameters, and issues the new parameters to all clients, starting the next round of training.

[0067] Repeat the above local training, parameter uploading and parameter aggregation process for 20 rounds of federated learning training. During the training process, the federated learning coordination end monitors the performance indicators of the global model, such as evaluating the accuracy and F1 score of the task classifier on a reserved validation set, and the discrimination accuracy of the domain discriminator. With the increase of training rounds, the performance of the task classifier gradually improves, and reaches a stable state around the 10th round, with an accuracy of 85% and an F1 score of 0.82 on the validation set. The discrimination accuracy of the domain discriminator is high in the initial stage, and gradually decreases with the adversarial training, and finally stabilizes at about 50%, close to the level of random guessing, which indicates that the embedding generator G has successfully learned the domain-invariant representation, making the domain discriminator unable to distinguish the source domain of the features.

[0068] After the training converges, the knowledge graph construction phase begins. Each client uses the trained embedding generator G to generate embedding vectors for all the knowledge fragments locally. The public domain client generates embedding vectors for 100,000 entities, and the private domain client generates embedding vectors for 5,000 entities. The clients upload these embedding vectors and their corresponding entity identifiers and relationship information to the federated learning coordination end or a dedicated knowledge graph construction server. When uploading, only embedding vectors and high-level semantic information are uploaded, and the original text content is not uploaded, thereby protecting data privacy. The knowledge graph construction server aligns entities in the global embedding space. The similarity between any two entity embedding vectors is calculated using cosine similarity. If the similarity is greater than the threshold value 0.9, the two entities are considered to be synonymous or related, and further alignment is required. Through clustering algorithms, entities with high similarity are clustered into a group. For example, the embedding vectors of "deeplearning" in the public domain and "deep learning" in the private domain are very close and are clustered into the same group. The server selects "deeplearning" as the standard name, records "deep learning" as the Chinese alias, and records in the bloodline tracking module that the two entities come from the public domain and the private domain respectively, the alignment time, and the alignment algorithm.

[0069] After all entities are aligned, the relationship information is integrated to construct a global knowledge graph. The triplets uploaded by each client are de-duplicated and fused. For multiple relationships between the same pair of entities, the most reliable relationship is selected according to the confidence, or multiple relationships are retained to form multi-element relationships. The final knowledge graph contains 50,000 unified entity nodes and 300,000 relationship edges, including technical terms, research methods, application fields, researchers, institutions, and other types of entities in public literature and internal reports, and "applied to", "belongs to", "cited", "cooperates", and other semantic relationships. The construction results of the knowledge graph are stored in a graph database, such as Neo4j, supporting efficient graph query and analysis.

[0070] The bloodline tracking module records the complete construction process. For each entity in the knowledge graph, the source domain, the original data reference, the label generation method used, the embedding vector generation model version, and other entities involved in the alignment process are recorded. For each relationship, the source client, the extraction method, and the confidence score are recorded. These bloodline metadata are organized into a separate metadata graph, which coexists with the content graph of the knowledge graph. Users can trace the source of any entity or relationship through the query interface. For example, querying the "deeplearning" entity, it can be seen that it comes from the public domain client, is extracted from the abstract of a certain paper by the BERT model, is aligned with the "deep learning" entity in the private domain, the alignment is based on the high similarity of embedding vectors, and the alignment occurs after the 15th training.

[0071] In the application stage, the constructed knowledge graph is used for various downstream tasks. In technical intelligence analysis, researchers can query the research status, main research institutions and key researchers of a specific technology through a graph query language such as Cypher, for example, querying "application of deep learning in medical imaging", returning related papers, internal projects, research teams and technology routes. In scientific research project management, project managers can use the knowledge graph for project planning and resource allocation, identify existing research results and potential partners related to new projects. In expert recommendation, the system recommends experts and scholars in related fields to users according to the information of the person entities and their research fields, published papers, etc. in the knowledge graph, and promotes academic exchange and cooperation. In technology evolution monitoring, by analyzing the changes of the knowledge graph over time, emerging technology hotspots and research trends are identified to provide data support for scientific research decision-making.

[0072] It can be seen from the embodiment that the method can effectively solve the privacy protection and semantic fusion of cross-domain data in practical application, and construct a high-quality cross-domain knowledge graph. Compared with the traditional centralized processing method, the method protects the privacy of internal data and meets the requirements of data security regulations. Compared with the traditional federated learning method, the method improves the domain generalization ability of the model through adversarial training and semantic alignment, so that data from different domains can be fused in a unified semantic space. Compared with existing knowledge graph construction technologies, the method can process semantically heterogeneous cross-domain data, automatically align entities and fuse relationships, and provide complete data bloodline tracking, enhancing the credibility and explainability of the knowledge graph. The successful application of the embodiment verifies the effectiveness and practicality of the technical solution of the application, providing an innovative solution for cross-domain knowledge graph construction.

[0073] The above only describes the preferred embodiments of the present application and is not intended to limit the protection scope of the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A multi-source data and knowledge graph construction method based on federated learning and adversarial training, characterized in that , comprising the following steps: S1, label generation and knowledge fragment construction: each client adopts a double-path label generation strategy to process local data, the double path includes a keyword-driven path and a path based on a natural language processing model, and generates structured knowledge fragments, which are represented in the form of triples (entity 1, relationship, entity 2) or (entity, attribute, attribute value); S2, federated learning initialization: the federated learning coordination end initializes the parameters of the embedding generator G, the task classifier C and the domain discriminator D, and distributes the initial parameters to each client, the embedding generator G is used to map the input data x to the latent feature space to generate an embedding vector z=G(x), and the domain discriminator D is connected to the embedding generator G through a gradient reversal layer; S3, local adversarial training: each client trains the model using local data, updates the model parameters by optimizing a joint loss function, which includes task loss , adversarial loss and semantic alignment loss three parts, the total loss function of the embedding generator G is , where λ and β are hyperparameters; in each training batch, first fix the parameters of the embedding generator G and the task classifier C, update the domain discriminator D; then fix the parameters of the domain discriminator D, update the embedding generator G and the task classifier C S4, parameter aggregation: after each client completes local training, the parameter updates of the embedding generator G and the task classifier C are calculated and uploaded, the federated learning coordination end uses the federated average algorithm to aggregate the collected parameter updates by weighting, the aggregation formula is , wherein is the data volume of the i-th client, is the total data volume, and the aggregated global model parameters are distributed to each client; S5, collaborative construction of knowledge graph: repeating steps S3 to S4 until the model converges, using the trained embedding generator G to generate domain-invariant embedding vectors for the knowledge fragments of each client, aligning and clustering similar entities in the global embedding space by calculating the similarity of embedding vectors, merging similar entities into unified entities, and integrating the relationship information of each client to construct a cross-domain knowledge graph; S6, bloodline tracking: record the data source, processing path, model version and fusion process for each entity and relationship in the knowledge graph, the data source includes domain identifier and client identifier, the processing path includes label generation strategy and generation time, and the fusion process includes aligned entities and alignment algorithm.

2. The method of claim 1, wherein , the task loss in step S3 comprising a label prediction loss and a knowledge graph construction loss , wherein: Label prediction loss where N is the number of samples in a batch, K is the total number of label classes, is the true label, is the predicted probability; Knowledge graph construction loss , wherein M is the number of triples, and are the subject entity and the object entity, respectively, is the relationship; task loss where γ is a balancing coefficient.

3. The method of claim 1, wherein The adversarial loss L_adv in step S3 is defined as: , wherein and are data distributions for public and private domains, respectively; The gradient reversal layer is used to realize adversarial training, which is an identity mapping during forward propagation and transmits the gradient multiplied by a negative number from the domain discriminator D back to the embedding generator G during backward propagation.

4. The method of claim 1, wherein , the semantic alignment loss in step S3 InfoNCE loss with contrastive learning, for cross-domain positive sample pairs The semantic alignment loss is defined as: , where P is the number of positive sample pairs, and is the embedding representation, sim(·, ·) is the similarity function, and τ is the temperature coefficient.

5. The method of claim 1, wherein The specific implementation of the double-path label generation strategy in step S1 is: The keyword-driven path uses the TF-IDF algorithm to extract keywords and key phrases from the text as entity labels; The path based on the natural language processing model uses a pre-trained language model to perform named entity recognition and relationship extraction on the text to generate structured knowledge fragments.

6. The method of claim 1, wherein , in the step S4, the client adds differential privacy noise to the parameter update before uploading the parameter update, and the parameter update after adding the noise is , wherein S is the sensitivity of the parameter update, ε is the privacy budget, and δ is the failure probability of the privacy loss.

7. The method of claim 1, wherein The specific steps of entity alignment in step S5 are: calculating the cosine similarity between any two entity embedding vectors, identifying two entities as synonymous entities when the similarity is greater than a preset threshold, using a clustering algorithm to group entities with high similarity into a group, selecting a standard entity name as a representative, and recording the original source information of each entity.

8. A multi-source data and knowledge graph construction system based on federated learning and adversarial training, characterized in that It comprises: a federated learning coordination end, used for initializing global model parameters, distributing model parameters to each client, collecting parameter updates uploaded by each client, using a federated averaging algorithm for parameter aggregation, and distributing the aggregated model parameters to each client; a plurality of clients, each client comprising: an embedding generator G, used to receive local data and output high-dimensional semantic embedding vectors, realizing a mapping function z=G(x); a domain discriminator D, which is a binary classification neural network, used to receive the vector z generated by the embedding generator G and judge its source domain, the domain discriminator D is connected to the embedding generator G through a gradient reversal layer; The task classifier C is a multi-label classification neural network, configured to receive the vector z generated by the embedding generator G and output a prediction result of multiple labels; The label generation module is configured to generate structured knowledge fragments by using a double-path label generation strategy, the double-path including a keyword-driven path and a path based on a natural language processing model; The local training module is configured to train the embedding generator G, the domain discriminator D and the task classifier C using local data, and update model parameters by optimizing a joint loss function including a task loss, an adversarial loss and a semantic alignment loss; The knowledge graph construction module is configured to generate embedding vectors for knowledge fragments of each client by using the trained embedding generator G, perform entity alignment and clustering in a global embedding space, and construct a cross-domain knowledge graph; The bloodline tracking module is configured to record data sources, processing paths, model versions and fusion processes of each entity and relationship in the knowledge graph.

9. The system of claim 8, wherein The embedding generator G adopts a pre-trained language model based on a Transformer architecture, and is fine-tuned in a federated learning process to achieve domain invariance; the domain discriminator D is composed of several fully connected layers, and outputs domain classification probabilities through a sigmoid activation function at the end; the task classifier C is composed of several fully connected layers, and outputs independent probability predictions for each label through a sigmoid activation function at the end.

10. The system of claim 8, wherein The information recorded by the bloodline tracking module includes: source domain identifiers, client identifiers, label generation strategies, model versions, generation times, other entities participating in alignment, alignment algorithms and alignment times of entities; and source clients, extraction methods and confidence scores of relationships.

Images