Multi-source heterogeneous big data fusion and reasoning method and system based on knowledge graph

Abstract

The invention relates to the technical field of big data analysis and artificial intelligence, in particular to a multi-source heterogeneous big data fusion and reasoning method and system based on a knowledge graph, and the method comprises the following steps: constructing mapping interfaces of various data sources, collecting and storing entities in the data sources, and constructing nodes in the knowledge graph; performing representation learning on a graph structure of the knowledge graph based on a graph neural network, extracting an embedded vector of a node, and extracting a fusion node through the embedded vector and a mapping interface; performing knowledge completion on the fusion nodes by using a neural network based on symbolic logic; generating candidate reasoning paths by adopting a depth-first search algorithm based on the knowledge graph; reasoning is carried out based on an attention mechanism and path similarity, candidate reasoning paths are screened and sorted to obtain an optimal reasoning path, a graph neural network is adopted to carry out representation learning of a knowledge graph, symbolic logic and the neural network are combined, the interpretability of the reasoning process is guaranteed, and the learning ability is high.

Description

Technical Field

[0001] The present invention relates to the field of big data analysis and artificial intelligence technology, and in particular to a multi-source heterogeneous big data fusion and reasoning method and system based on a knowledge graph. Background Art

[0002] With the advent of the big data era, we are facing an unprecedented information explosion. All walks of life are generating massive amounts of data, which come from various sources and in various forms, including structured database records, semi-structured XML files, and unstructured text documents. How to effectively integrate and utilize these heterogeneous data has become a major challenge in the current information technology field.

[0003] In recent years, knowledge graphs, as powerful knowledge representation and reasoning tools, have shown great potential in processing complex data relationships. They can intuitively represent complex relationships between entities in the form of graph structures, providing an ideal platform for data fusion and knowledge reasoning. However, traditional knowledge graph construction methods often have difficulty in coping with the fusion of multi-source heterogeneous data.

[0004] At present, the commonly used method in the industry is to convert various data sources into a unified format before constructing the knowledge graph. Although this method is simple and direct, it often causes a lot of valuable information to be lost during the conversion process. Especially for data that is difficult to directly map to a predefined pattern, the potential relationships contained therein are likely to be overlooked. In addition, this method is also difficult to handle dynamically changing data sources, and it is impossible to update the knowledge graph in a timely manner to reflect the latest information.

[0005] On the other hand, in terms of knowledge reasoning, most existing methods rely on predefined reasoning rules or statistical learning models. These methods perform well when dealing with reasoning tasks with known patterns, but they are often unable to cope with complex, cross-domain reasoning problems. They have difficulty capturing the deep semantic associations between data, and cannot effectively utilize the rich information contained in the graph structure.

[0006] In addition, existing reasoning methods are usually designed for specific fields or tasks, lacking versatility and scalability. When faced with new data sources or new reasoning tasks, these methods often require a lot of manual intervention and adjustment, making it difficult to adapt to rapidly changing business needs.

[0007] In the face of these challenges, we urgently need new methods that can effectively integrate multi-source heterogeneous data, build a comprehensive and dynamic knowledge graph, and realize intelligent knowledge reasoning on this basis. This method should be able to automatically process different types of data sources, capture the potential correlation between data, and be able to self-improve and update as data continues to accumulate. At the same time, it should also have strong reasoning capabilities, be able to handle complex, cross-domain reasoning tasks, and provide strong support for various intelligent applications. Summary of the invention

[0008] The present invention proposes an innovative solution to the above problems. It proposes a multi-source heterogeneous big data fusion and reasoning method and system based on knowledge graph, aiming to achieve seamless fusion of multi-source heterogeneous data, dynamic construction and update of knowledge graph, and intelligent reasoning based on graph structure. This method cleverly combines advanced technologies such as graph neural network, neural symbolic reasoning, attention mechanism, etc., and has achieved breakthrough progress in data fusion, knowledge representation and reasoning ability.

[0009] The present invention proposes a multi-source heterogeneous big data fusion and reasoning method based on knowledge graph, which includes the following steps:

[0010] Build mapping interfaces for multiple data sources, collect and store entities in data sources, and build nodes in knowledge graphs;

[0011] Subsequently, representation learning is performed on the graph structure of the knowledge graph based on a graph neural network, embedding vectors of nodes are extracted, and fusion nodes are extracted through the embedding vectors and the mapping interface;

[0012] Next, using a neural network based on symbolic logic to complete the knowledge of the fusion node;

[0013] Then, a depth-first search algorithm is used to generate candidate reasoning paths based on the knowledge graph;

[0014] Finally, based on the attention mechanism and path similarity, reasoning is performed to screen and sort the candidate reasoning paths to obtain the optimal reasoning path.

[0015] Preferably, the steps of constructing mapping interfaces for multiple data sources include:

[0016] First, multiple data sources are obtained through web crawlers or API interfaces. Each data source contains different types of entities, including people, geographic locations, organizations, events, and relationships.

[0017] Secondly, the data source is cleaned and preprocessed to remove redundant and erroneous data;

[0018] Then, a mapping rule between the data source and the entity is defined to map the data from different data sources into a unified knowledge graph format;

[0019] Next, the mapped data is stored in the knowledge graph to construct a data model of the knowledge graph, wherein the data model of the knowledge graph is composed of a set of nodes, where the nodes represent different entities, such as people, geographic locations, and events, and the relationships between the nodes represent different relationships, such as "colleague relationship", "location", and "organizational structure";

[0020] Finally, natural language processing technology is used to extract information about the entities and the relationships, and the knowledge graph is updated.

[0021] Preferably, the steps of performing representation learning based on graph neural network include:

[0022] First, for the graph-structured knowledge graph, each node is represented by an embedding vector. This embedding vector is a vector representation in the semantic space, which contains the information of the node in the knowledge graph;

[0023] Secondly, the similarity between nodes is calculated to characterize the relevance of nodes. The embedding learning model based on graph neural network learns the embedding vector of nodes through training data, and makes nodes with similar structures or similar semantics have similar or close embedding vectors.

[0024] Then, the embedding vector of each node is learned through the graph neural network to convert the knowledge graph of the graph structure into a semantic vector;

[0025] Finally, fusion nodes are extracted through the embedding vector and the mapping interface. For multiple data sources with similar nodes, the nodes are fused according to the mapping rules and the embedding vector to obtain a fusion node that integrates information from multiple data sources. The fusion node contains information from multiple data sources, including relationships, attributes, and text descriptions.

[0026] Preferably, the step of using a neural network based on symbolic logic to complete knowledge includes:

[0027] First, the attribute information and text description of the fused node are input into a neural network based on symbolic logic to extract symbolic logic and feature vectors. In the neural network, the attribute information of the fused node is converted into symbolic logic, which includes the attribute value of the node itself, the relationship between nodes and constraint conditions. At the same time, the text description of the fused node is embedded through the attention mechanism to obtain a text feature vector.

[0028] Secondly, the symbolic logic and the feature vector are input into a symbolic logic-based neural network for reasoning, and the network adopts a convolutional neural network or a recurrent neural network model to combine the symbolic logic and the feature vector for knowledge completion;

[0029] Then, the inference network outputs a probability distribution representing the credibility of different constraints in the symbolic logic;

[0030] Finally, according to the probability distribution, the constraints in the symbolic logic are screened to generate new knowledge triples. For example, when it is inferred that the relationship between two nodes is "colleagues", the probability distribution will output a credibility representing the relationship. If the credibility is higher than the set threshold, the two nodes and the "colleague" relationship are taken as a new knowledge triple and completed in the knowledge graph.

[0031] Preferably, the step of generating a candidate reasoning path comprises:

[0032] Firstly, based on the knowledge graph and the fusion node, a depth-first search algorithm is used to traverse the nodes and edges in the knowledge graph to generate all possible reasoning paths, wherein the reasoning path is a set of a series of knowledge triples;

[0033] Secondly, the nodes in each reasoning path are converted into embedded vectors, that is, into a series of vectors in the semantic space of the knowledge graph, to obtain an embedded vector set of a reasoning path;

[0034] Finally, the embedding vector set of the reasoning path is input into the attention mechanism to calculate the similarity of the reasoning path, that is, the cosine similarity of the path embedding vector.

[0035] Preferably, the steps of reasoning based on the attention mechanism and path similarity include:

[0036] First, all possible reasoning paths are sorted according to their similarity, and the reasoning path with the highest similarity, that is, the reasoning path with the highest cosine similarity of the path embedding vector, is considered to be the optimal reasoning path;

[0037] Secondly, based on the attention mechanism, according to the reasoning path and the reasoning target, that is, the relationship specified in the query statement, the attention weight is calculated. The attention weight is calculated based on the distance between different nodes of the reasoning path and the relationship specified in the query statement and the structural related information. The attention weight will be given to different reasoning paths, that is, the similarity is weighted according to the distance between different nodes of the path and the relationship specified in the query statement and the structural related information;

[0038] Finally, the attention weight is combined with the similarity to calculate the score of the reasoning path, so as to screen and sort the reasoning paths and obtain the optimal reasoning path.

[0039] A multi-source heterogeneous big data fusion and reasoning system based on a knowledge graph for executing the method includes a data source mapping interface module, a graph neural network module, a neural symbolic reasoning module, and a reasoning path screening and sorting module;

[0040] The data source mapping interface module is used to obtain multiple data sources through a web crawler or an API interface, each data source contains different types of entities, and map the data from different data sources to a unified knowledge graph format by defining mapping rules between data sources and entities;

[0041] The graph neural network module is used to extract information about entities and relationships from data in the data source through natural language processing technology, obtain the embedding vector of each node based on graph neural network learning, and use the embedded vector knowledge graph to extract features to achieve node fusion;

[0042] The neural symbolic reasoning module is used to represent and learn the attribute information and text description information of the fused nodes, that is, the features and knowledge in multiple data sources, perform knowledge reasoning, and realize knowledge completion of the knowledge graph;

[0043] The inference path screening and sorting module is used to screen and sort the inference paths, and select the optimal inference path from the candidate inference paths. The optimal inference path is the inference path with the highest similarity and is combined with the attention mechanism to perform score calculation.

[0044] Preferably, the data source mapping interface module includes a web crawler crawling unit, an API interface access unit, a data cleaning unit, a data storage unit and a human-computer interaction unit;

[0045] The web crawler unit is used to collect data from multiple data sources through a web crawler, wherein the web crawler is a crawler that supports a data source interface and can support automatic acquisition of data from various websites, databases and other data sources;

[0046] The API interface access unit is used to access various data sources through the API interface to obtain data;

[0047] The data cleaning unit is used to remove redundant and erroneous data by cleaning and preprocessing the collected data;

[0048] The data storage unit is used to construct a data model of the knowledge graph by storing the mapped data in the knowledge graph;

[0049] The human-computer interaction unit is used to display the knowledge graph through a user interface and receive user queries and instructions.

[0050] Preferably, the workflow of the graph neural network module is:

[0051] First, obtain data from multiple data sources through web crawlers or API interfaces;

[0052] Secondly, by defining the mapping rules between data sources and entities, the data from different data sources are mapped to a unified knowledge graph format;

[0053] Then, the data embedding learning model learns the embedding vectors of nodes through training data and converts the knowledge graph of the graph structure into semantic vectors;

[0054] Finally, the fused nodes are extracted through the embedding vector and mapping interface.

[0055] Preferably, the workflow of the neural symbolic reasoning module is:

[0056] First, the attribute information, text description information and knowledge of the node are represented as symbolic logic, that is, the attribute information and text description information of the fused node are converted into symbolic logic and feature vectors through the attention model;

[0057] Secondly, the symbolic logic and the feature vector are input into a neural network based on symbolic logic to infer and generate new knowledge triples, that is, new knowledge triples are inferred based on the constraints contained in the symbolic logic and the information in the feature vector to achieve knowledge completion of the knowledge graph;

[0058] Finally, the knowledge in the knowledge graph is updated according to the generated knowledge triples to achieve knowledge updating.

[0059] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0060] The present invention can effectively process various types of data sources, including structured, semi-structured and unstructured data, by designing a flexible data source mapping interface. It uses graph neural networks to perform representation learning of knowledge graphs, and can make full use of the topological information of the graph structure to capture the complex relationships between entities. In terms of knowledge reasoning, the present invention innovatively combines symbolic logic and neural networks, which not only ensures the interpretability of the reasoning process, but also has powerful learning capabilities. In addition, by introducing the attention mechanism, the present invention can more accurately select relevant information and improve the accuracy and efficiency of reasoning.

[0061] It is worth mentioning that the present invention performs well in resolving algorithmic mechanism contradictions. For example, it cleverly balances the comprehensiveness of data fusion and computational efficiency, and through a layered data processing strategy, it not only ensures the integrity of information but also avoids the waste of computing resources. In terms of knowledge reasoning, the present invention effectively combines the accuracy of symbolic reasoning and the generalization ability of neural networks through neural symbolic integrated learning, overcoming the limitations of a single method.

[0062] In addition, the various components of the present invention show significant synergistic effects. The data fusion module provides a rich and accurate information basis for the construction of the knowledge graph, and the embedded representation learned by the graph neural network provides powerful feature support for subsequent reasoning tasks. The introduction of the attention mechanism further enhances the adaptability and accuracy of the entire system, making the reasoning results more reliable and convincing.

[0063] In general, the present invention not only solves the technical problem of multi-source heterogeneous data fusion, but also greatly improves the reasoning ability based on knowledge graph. It opens up a new way to build intelligent and adaptive knowledge systems, and has broad application prospects in many fields such as intelligent customer service, financial risk assessment, and smart city management. Through the present invention, we can more fully explore the value of big data and provide strong support for the intelligent transformation of all walks of life. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 It is a logic block diagram of the whole system of the present invention.

[0065] Figure 2 This is a logic block diagram of the data source mapping interface module of the present invention.

[0066] Figure 3 A logic block diagram of the method steps of the present invention.

[0067] Figure 4 It is a logical block diagram of the graphless neural network module of the present invention.

[0068] Figure 5 It is a logical block diagram of the neural symbolic reasoning module of the present invention.

[0069] Figure 6 A logical block diagram for generating and screening the reasoning paths of the present invention. DETAILED DESCRIPTION

[0070] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects thereof are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0071] Unless defined otherwise, 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 belongs.

[0072] See also Figure 1-6 The present invention relates to a multi-source heterogeneous big data fusion and reasoning method and a system based on a knowledge graph, particularly in the field of big data analysis and artificial intelligence, for solving the fusion and knowledge reasoning problems of multi-source heterogeneous data.

[0073] In a specific embodiment of the present invention, the present invention proposes a multi-source heterogeneous big data fusion and reasoning method based on knowledge graph, which includes the following key steps:

[0074] First, we build mapping interfaces for multiple data sources, collect and store entities in the data sources, and build nodes in the knowledge graph. In this step, we can effectively process heterogeneous data from different sources, such as structured databases, semi-structured XML files, and unstructured text documents, by designing flexible mapping interfaces. For example, in the financial field, we can simultaneously process real-time data from stock trading systems, text information from company financial reports, and public opinion data on social media.

[0075] Subsequently, the graph structure of the knowledge graph is represented and learned based on the graph neural network, the embedding vectors of the nodes are extracted, and the fusion nodes are extracted through the embedding vectors and the mapping interface. The core of this step is to use the powerful representation ability of the graph neural network to transform the complex knowledge graph structure into a low-dimensional vector representation. For example, we can use advanced graph neural network models such as GraphSAGE or GAT to perform efficient representation learning on large-scale knowledge graphs containing millions of nodes.

[0076] Next, the fusion node is completed using a neural network based on symbolic logic. This step combines the interpretability of symbolic logic with the learning ability of neural networks, and can effectively infer implicit knowledge relationships. For example, in a medical diagnosis system, we can infer potential disease risks based on known symptoms and disease relationships.

[0077] Then, a depth-first search algorithm is used based on the knowledge graph to generate candidate reasoning paths. This step generates possible reasoning paths by systematically exploring the structure of the knowledge graph. For example, in a recommendation system, we can discover potential connections between user interests in this way.

[0078] Finally, based on the attention mechanism and path similarity, the candidate reasoning paths are screened and sorted to obtain the optimal reasoning path. This step can better capture the key information in the path and improve the accuracy of reasoning by introducing the attention mechanism. For example, in an intelligent customer service system, we can quickly find the most relevant question-solving path in this way.

[0079] In a specific embodiment of the present invention, the specific steps of constructing a mapping interface for multiple data sources are as follows:

[0080] First, multiple data sources are obtained through web crawlers or API interfaces. Each data source contains different types of entities, including people, geographic locations, organizations, events, and relationships. The key to this step is the diversity and comprehensiveness of the data sources. For example, in a smart city project, we may need to obtain real-time data from the traffic system, weather forecasts from the meteorological department, and citizen feedback on social media at the same time.

[0081] Secondly, the data source is cleaned and preprocessed to remove redundant and erroneous data. This step is crucial to ensure the accuracy of subsequent analysis. We can use some efficient data cleaning algorithms, such as outlier detection, missing value filling, etc. For example, when processing sensor data, we may need to remove abnormal readings caused by equipment failure.

[0082] Then, we define the mapping rules between the data sources and the entities, and map the data from different data sources to a unified knowledge graph format. This step is the key to achieving heterogeneous data fusion. We can design flexible mapping rules, such as using ontology matching algorithms to automatically discover the semantic correspondence between different data sources.

[0083] Next, the mapped data is stored in the knowledge graph to build a data model of the knowledge graph. In this step, we need to select a suitable knowledge graph storage and query system, such as Neo4j or JanusGraph, to support efficient storage and query of large-scale knowledge graphs.

[0084] Finally, natural language processing technology is used to extract information about the entities and relationships, and the knowledge graph is updated. This step can extract structured knowledge from unstructured text and continuously enrich and update the knowledge graph. For example, we can use technologies such as named entity recognition and relationship extraction to extract cooperative relationships between companies from news articles and update the business knowledge graph in real time.

[0085] In a specific embodiment of the present invention, the steps of performing representation learning based on a graph neural network are as follows:

[0086] First, for the graph structured knowledge graph, each node is represented by an embedding vector, which is a vector representation in the semantic space and contains the information of the node in the knowledge graph. For example, in a knowledge graph containing entities such as movies, actors, and directors, each entity can be represented by a 300-dimensional vector that captures the semantic features of the entity.

[0087] Secondly, the similarity between nodes is calculated to characterize the relevance of nodes. The embedding learning model based on graph neural network learns the embedding vector of nodes through training data, and makes nodes with similar structures or similar semantics have similar or close embedding vectors. The core of this step is to select a suitable graph neural network model and similarity measurement method. For example, we can use cosine similarity to measure the similarity between vectors, and the threshold can be set to 0.8. This value is an empirical value based on a large number of experiments, which can effectively filter out noise while retaining relevance.

[0088] Then, the graph neural network is used to learn the embedding vector of each node and transform the graph structured knowledge graph into a semantic vector. In this step, we can use advanced graph neural network models such as GraphSAGE or GAT. For example, for a knowledge graph containing 1 million nodes, we can set the number of GraphSAGE layers to 3 and the sampling size to 10-25-10, which can achieve a good balance between computational efficiency and representation ability.

[0089] Finally, the fusion node is extracted through the embedding vector and the mapping interface. For multiple data sources with similar nodes, the nodes are fused according to the mapping rules and the embedding vector to obtain a fusion node that integrates the information of multiple data sources. The fusion node contains information from multiple data sources, including relationships, attributes, and text descriptions. The key to this step is how to effectively fuse information from different data sources. For example, we can use a weighted average method to fuse vector representations from different sources, and the weights can be set based on the reliability of the data source.

[0090] In a specific embodiment of the present invention, the steps of using a neural network based on symbolic logic to complete knowledge are as follows:

[0091] First, the attribute information and text description of the fusion node are input into a neural network based on symbolic logic to extract symbolic logic and feature vectors. In this step, we need to design a neural network model that can process both structured attributes and unstructured text. For example, we can use a pre-trained language model such as BERT to process text descriptions, a multi-layer perceptron to process attribute information, and then fuse the outputs of the two.

[0092] Secondly, the symbolic logic and the feature vector are input into a symbolic logic-based neural network for reasoning. The network uses a convolutional neural network or a recurrent neural network model to combine the symbolic logic and the feature vector for knowledge completion. The core of this step is how to effectively combine symbolic logic and neural networks. For example, we can use the method of neural symbolic ensemble learning to embed symbolic logic rules into the structure of the neural network, thereby achieving explainable knowledge reasoning.

[0093] Then, the inference network outputs a probability distribution, which indicates the credibility of different constraints in the symbolic logic. This step provides a reliability assessment for subsequent knowledge completion. For example, in a medical diagnosis system, we can evaluate the likelihood of different diagnosis results based on this probability distribution.

[0094] Finally, according to the probability distribution, the constraints in the symbolic logic are screened to generate new knowledge triples. For example, when the relationship between two nodes is inferred to be "colleagues", the probability distribution will output a credibility representing the relationship. If the credibility is higher than the set threshold, the two nodes and the "colleague" relationship are taken as a new knowledge triple and completed in the knowledge graph. In practical applications, we can set the threshold according to the requirements of specific tasks. For example, in a system that requires high accuracy, we may set the threshold to 0.9, while in a system that focuses more on recall, we may set the threshold to 0.7.

[0095] In a specific embodiment of the present invention, the steps of generating candidate reasoning paths are as follows:

[0096] First, based on the knowledge graph and the fusion node, a depth-first search algorithm is used to traverse the nodes and edges in the knowledge graph to generate all possible reasoning paths, where the reasoning path is a set of knowledge triples. The key to this step is how to efficiently traverse a large-scale graph. For example, we can set the maximum search depth to 5, which is an empirical value based on a large number of experiments, and can control the computational complexity while ensuring the quality of the reasoning path.

[0097] Secondly, the nodes in each reasoning path are converted into embedded vectors, that is, into a series of vectors in the semantic space of the knowledge graph, to obtain a set of embedded vectors for a reasoning path. This step utilizes the node embedding representation learned previously, and can represent the path as a series of semantic vectors.

[0098] Finally, the embedding vector set of the reasoning path is input into the attention mechanism to calculate the similarity of the reasoning path, that is, the cosine similarity of the path embedding vector. This step can better capture the key information in the path by introducing the attention mechanism. For example, we can use the self-attention mechanism to calculate the importance weights of different nodes in the path, thereby obtaining a more accurate path representation.

[0099] In a specific embodiment of the present invention, the steps of reasoning based on the attention mechanism and path similarity are as follows:

[0100] First, all possible reasoning paths are sorted according to their similarity. The reasoning path with the highest similarity, that is, the reasoning path with the highest cosine similarity of the path embedding vector, is considered the optimal reasoning path. The key to this step is how to efficiently sort a large number of reasoning paths. For example, we can use the heap sort algorithm to quickly select the Top-K most similar paths.

[0101] Secondly, based on the attention mechanism, the attention weight is calculated according to the reasoning path and the reasoning target, that is, the relationship specified in the query statement. The attention weight is calculated based on the distance, structure and other related information between the different nodes of the reasoning path and the relationship specified in the query statement. The attention weight will be assigned to different reasoning paths, that is, the similarity is weighted according to the distance, structure and other related information between the different nodes of the path and the relationship specified in the query statement. This step can better capture the correlation between the query target and the reasoning path by introducing the attention mechanism. For example, we can use a multi-head attention mechanism and set the number of heads to 8, so that correlation information can be captured from multiple angles.

[0102] Finally, the attention weight is combined with the similarity to calculate the score of the reasoning path, so as to screen and sort the reasoning path and obtain the optimal reasoning path. This step comprehensively considers the path similarity and attention weight, and can obtain more accurate reasoning results. For example, we can use a weighted sum method to combine similarity and attention weight, and the weight can be determined by cross-validation to determine the optimal value.

[0103] The present invention also proposes a multi-source heterogeneous big data fusion and reasoning system based on knowledge graph. The system includes the following key modules:

[0104] First, the data source mapping interface module 1 is used to obtain multiple data sources through web crawlers or API interfaces. Each data source contains different types of entities. By defining the mapping rules between data sources and entities, the data from different data sources are mapped to a unified knowledge graph format. This module is the data entry for the entire system, and its flexibility and scalability directly affect the system's ability to process heterogeneous data. For example, in a smart city project, this module can simultaneously process real-time data streams from the traffic system, structured databases from the meteorological department, and unstructured text data from social media.

[0105] Secondly, the graph neural network module 2 is used to extract information about entities and relationships from the data source through natural language processing technology, obtain the embedding vector of each node based on graph neural network learning, and use the embedded vector knowledge graph to extract features and realize node fusion. This module is the core of the system and is responsible for converting complex knowledge graph structures into vector representations that are easy for computers to process. For example, we can use the GraphSAGE model and set a 3-layer network structure with a sampling size of 10-25-10 for each layer. This setting performs well on large-scale knowledge graphs.

[0106] Then, the neural symbolic reasoning module 3 is used to represent and learn the attribute information and text description information of the fusion node, that is, the features and knowledge in multiple data sources, perform knowledge reasoning, and realize the knowledge completion of the knowledge graph. This module combines the interpretability of symbolic logic and the learning ability of neural networks, and can effectively infer implicit knowledge relationships. For example, in a financial risk assessment system, this module can infer potential financial risks based on various attribute information of the company and related news reports.

[0107] Finally, the reasoning path screening and sorting module 4 is used to screen and sort the reasoning paths, and select the optimal reasoning path from the candidate reasoning paths. The optimal reasoning path is the reasoning path with the highest similarity and the score is calculated through the attention mechanism. This module is the key to the system output results, which determines the quality of the final reasoning results. For example, in an intelligent question-answering system, this module can select the most relevant and accurate answer path from many possible answer paths.

[0108] In a specific embodiment of the present invention, the structure of the data source mapping interface module 1 is as follows:

[0109] First, the web crawler crawling unit 11 is used to collect data from multiple data sources through a web crawler, wherein the web crawler is a crawler that supports a data source interface and can support automatic acquisition of data from various websites, databases, and other data sources. The design of this unit needs to take into account the diversity and dynamics of the web page structure. For example, we can use the Scrapy framework to build a flexible crawler system and combine it with tools such as Selenium to process dynamically loaded web page content.

[0110] Secondly, the API interface access unit 12 is used to access various data sources through the API interface and obtain data. This unit needs to be able to handle various API protocols and data formats. For example, we can use RESTful API to access structured data and use GraphQL to query graph structured data.

[0111] Then, the data cleaning unit 13 is used to remove redundant and erroneous data by cleaning and preprocessing the collected data. This unit is crucial to ensure the accuracy of subsequent analysis. For example, we can use Python's pandas library to perform efficient data cleaning, such as processing missing values, removing duplicates, and standardizing data formats.

[0112] Next, the data storage unit 14 is used to build a data model of the knowledge graph by storing the mapped data in the knowledge graph. This unit needs to select a suitable storage system to support efficient storage and query of large-scale knowledge graphs. For example, we can use Neo4j graph database to store knowledge graphs, which supports efficient graph structure storage and query.

[0113] Finally, the human-computer interaction unit 15 is used to display the knowledge graph through the user interface and receive user queries and instructions. This unit is a window for the system to interact with the user, and an intuitive and friendly interface needs to be designed. For example, we can use the D3.js library to realize the visualization of the knowledge graph and use natural language processing technology to support the user's natural language query.

[0114] In a specific embodiment of the present invention, the workflow of the graph neural network module 2 is as follows:

[0115] First, data from multiple data sources are obtained through web crawlers or API interfaces, etc. This step needs to be closely coordinated with the data source mapping interface module 1 to ensure the comprehensiveness and timeliness of the data.

[0116] Secondly, by defining the mapping rules between data sources and entities, the data from different data sources are mapped to a unified knowledge graph format. This step is the key to achieving heterogeneous data fusion. For example, we can use ontology matching algorithms to automatically discover the semantic correspondence between different data sources.

[0117] Then, the embedding learning model of the data learns the embedding vectors of the nodes through the training data and converts the graph structure knowledge graph into semantic vectors. In this step, we can use advanced graph neural network models such as Graph SAGE or GAT. For example, for a knowledge graph containing millions of nodes, we can set the training batch size of GraphSAGE to 1024, the learning rate to 0.01, and the number of training rounds to 100. Such settings can achieve a good balance between training efficiency and model performance.

[0118] Finally, the fusion node is extracted through the embedding vector and mapping interface. The key to this step is how to effectively fuse information from different data sources. For example, we can use a weighted average method to fuse vector representations from different sources, and the weights can be set based on the reliability of the data source.

[0119] In a specific embodiment of the present invention, the workflow of the neural symbolic reasoning module 3 is as follows:

[0120] First, the attribute information, text description information and knowledge of the node are represented as symbolic logic, that is, the attribute information and text description information of the fused node are converted into symbolic logic and feature vectors through the attention model. This step requires the design of a model that can process both structured attributes and unstructured text. For example, we can use a pre-trained language model such as BERT to process text descriptions, use a multi-layer perceptron to process attribute information, and then fuse the outputs of the two.

[0121] Secondly, the symbolic logic and the feature vector are input into a neural network based on symbolic logic to infer and generate new knowledge triples, that is, new knowledge triples are inferred based on the constraints contained in the symbolic logic and the information in the feature vector to achieve knowledge completion of the knowledge graph. The core of this step lies in how to effectively combine symbolic logic and neural networks. For example, we can use the method of neural symbolic integrated learning to embed symbolic logic rules into the structure of the neural network, thereby achieving explainable knowledge reasoning.

[0122] Finally, the knowledge in the knowledge graph is updated according to the generated knowledge triples to achieve knowledge update. This step requires the design of an efficient knowledge graph update mechanism. For example, we can use the incremental update method to only add or modify the newly generated knowledge triples instead of reconstructing the entire knowledge graph, thereby improving the real-time performance of the system.

[0123] In a preferred embodiment of the present invention, we can apply the method and system to the field of intelligent customer service. In this scenario, the knowledge graph can contain multiple types of entities and relationships such as product information, common problems, user feedback, etc. The system can process user queries in real time and use the knowledge graph for reasoning to quickly find the most relevant answer path. For example, when a user asks "how to solve the problem of a mobile phone not turning on", the system can use the knowledge graph to reason, consider multiple aspects of information such as mobile phone models, common faults, solutions, etc., and give the most suitable answer path.

[0124] In another embodiment of the present invention, we can apply the method and system to the field of financial risk assessment. In this scenario, the knowledge graph can contain multi-source heterogeneous data such as company information, financial data, market dynamics, policies and regulations. The system can discover potential risk factors and associations by fusing and reasoning about these data. For example, the system can analyze a company's supply chain relationships, financial status, industry dynamics and other information to infer the financial risks that the company may face and provide support for investment decisions.

[0125] In general, the multi-source heterogeneous big data fusion and reasoning method and system based on knowledge graph proposed in the present invention realizes effective fusion and intelligent reasoning of multi-source heterogeneous big data by combining advanced technologies such as graph neural network, neural symbolic reasoning, attention mechanism, etc. The method and system have significant advantages in improving data utilization efficiency, enhancing knowledge discovery capabilities, supporting intelligent decision-making, etc., and have broad application prospects in many fields such as intelligent customer service, financial risk assessment, and smart city management.

[0126] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multi-source heterogeneous big data fusion and reasoning method based on knowledge graph, characterized by: The steps include: Build mapping interfaces for multiple data sources, collect and store entities in data sources, and build nodes in knowledge graphs; Subsequently, representation learning is performed on the graph structure of the knowledge graph based on a graph neural network, embedding vectors of nodes are extracted, and fusion nodes are extracted through the embedding vectors and the mapping interface; Next, using a neural network based on symbolic logic to complete the knowledge of the fusion node; Then, a depth-first search algorithm is used to generate candidate reasoning paths based on the knowledge graph; Finally, based on the attention mechanism and path similarity, reasoning is performed to screen and sort the candidate reasoning paths to obtain the optimal reasoning path.

2. The multi-source heterogeneous big data fusion and reasoning method based on knowledge graph according to claim 1 is characterized in that: The steps to build mapping interfaces for multiple data sources include: First, multiple data sources are obtained through web crawlers or API interfaces. Each data source contains different types of entities, including people, geographic locations, organizations, events, and relationships. Secondly, the data source is cleaned and preprocessed to remove redundant and erroneous data; Then, a mapping rule between the data source and the entity is defined to map the data from different data sources into a unified knowledge graph format; Next, the mapped data is stored in the knowledge graph to construct a data model of the knowledge graph, wherein the data model of the knowledge graph is composed of a set of nodes, where the nodes represent different entities, such as people, geographic locations, and events, and the relationships between the nodes represent different relationships, such as "colleague relationship", "location", and "organizational structure"; Finally, natural language processing technology is used to extract information about the entities and the relationships, and the knowledge graph is updated.

3. The multi-source heterogeneous big data fusion and reasoning method based on knowledge graph according to claim 1 is characterized in that: The steps of representation learning based on graph neural network include: First, for the graph-structured knowledge graph, each node is represented by an embedding vector. This embedding vector is a vector representation in the semantic space, which contains the information of the node in the knowledge graph; Secondly, the similarity between nodes is calculated to characterize the relevance of nodes. The embedding learning model based on graph neural network learns the embedding vector of nodes through training data, and makes nodes with similar structures or similar semantics have similar or close embedding vectors. Then, the embedding vector of each node is learned through the graph neural network to convert the knowledge graph of the graph structure into a semantic vector; Finally, fusion nodes are extracted through the embedding vector and the mapping interface. For multiple data sources with similar nodes, the nodes are fused according to the mapping rules and the embedding vector to obtain a fusion node that integrates information from multiple data sources. The fusion node contains information from multiple data sources, including relationships, attributes, and text descriptions.

4. The multi-source heterogeneous big data fusion and reasoning method based on knowledge graph according to claim 1 is characterized in that: The steps of using a neural network based on symbolic logic to complete knowledge include: First, the attribute information and text description of the fused node are input into a neural network based on symbolic logic to extract symbolic logic and feature vectors. In the neural network, the attribute information of the fused node is converted into symbolic logic, which includes the attribute value of the node itself, the relationship between nodes and constraint conditions. At the same time, the text description of the fused node is embedded through the attention mechanism to obtain a text feature vector. Secondly, the symbolic logic and the feature vector are input into a symbolic logic-based neural network for reasoning, and the network adopts a convolutional neural network or a recurrent neural network model to combine the symbolic logic and the feature vector for knowledge completion; Then, the inference network outputs a probability distribution representing the credibility of different constraints in the symbolic logic; Finally, according to the probability distribution, the constraints in the symbolic logic are screened to generate new knowledge triples. For example, when it is inferred that the relationship between two nodes is "colleagues", the probability distribution will output a credibility representing the relationship. If the credibility is higher than the set threshold, the two nodes and the "colleague" relationship are taken as a new knowledge triple and completed in the knowledge graph.

5. The multi-source heterogeneous big data fusion and reasoning method based on knowledge graph according to claim 1 is characterized in that: The steps to generate candidate inference paths include: Firstly, based on the knowledge graph and the fusion node, a depth-first search algorithm is used to traverse the nodes and edges in the knowledge graph to generate all possible reasoning paths, wherein the reasoning path is a set of a series of knowledge triples; Secondly, the nodes in each reasoning path are converted into embedded vectors, that is, into a series of vectors in the semantic space of the knowledge graph, to obtain an embedded vector set of a reasoning path; Finally, the embedding vector set of the reasoning path is input into the attention mechanism to calculate the similarity of the reasoning path, that is, the cosine similarity of the path embedding vector.

6. The multi-source heterogeneous big data fusion and reasoning method based on knowledge graph according to claim 1 is characterized in that: The steps of reasoning based on attention mechanism and path similarity include: First, all possible reasoning paths are sorted according to their similarity, and the reasoning path with the highest similarity, that is, the reasoning path with the highest cosine similarity of the path embedding vector, is considered to be the optimal reasoning path; Secondly, based on the attention mechanism, according to the reasoning path and the reasoning target, that is, the relationship specified in the query statement, the attention weight is calculated. The attention weight is calculated based on the distance between different nodes of the reasoning path and the relationship specified in the query statement and the structural related information. The attention weight will be given to different reasoning paths, that is, the similarity is weighted according to the distance between different nodes of the path and the relationship specified in the query statement and the structural related information; Finally, the attention weight is combined with the similarity to calculate the score of the reasoning path, so as to screen and sort the reasoning paths and obtain the optimal reasoning path.

7. A multi-source heterogeneous big data fusion and reasoning system based on a knowledge graph that executes the method according to any one of claims 1 to 6, characterized in that: It includes data source mapping interface module, graph neural network module, neural symbolic reasoning module and reasoning path screening and sorting module; The data source mapping interface module is used to obtain multiple data sources through a web crawler or an API interface, each data source contains different types of entities, and map the data from different data sources to a unified knowledge graph format by defining mapping rules between data sources and entities; The graph neural network module is used to extract information about entities and relationships from data in the data source through natural language processing technology, obtain the embedding vector of each node based on graph neural network learning, and use the embedded vector knowledge graph to extract features to achieve node fusion; The neural symbolic reasoning module is used to represent and learn the attribute information and text description information of the fused nodes, that is, the features and knowledge in multiple data sources, perform knowledge reasoning, and realize knowledge completion of the knowledge graph; The inference path screening and sorting module is used to screen and sort the inference paths, and select the optimal inference path from the candidate inference paths. The optimal inference path is the inference path with the highest similarity and is combined with the attention mechanism to perform score calculation.

8. The multi-source heterogeneous big data fusion and reasoning system based on knowledge graph according to claim 7 is characterized in that: The data source mapping interface module includes a web crawler crawling unit, an API interface access unit, a data cleaning unit, a data storage unit and a human-computer interaction unit; The web crawler unit is used to collect data from multiple data sources through a web crawler, wherein the web crawler is a crawler that supports a data source interface and can support automatic acquisition of data from various websites, databases and other data sources; The API interface access unit is used to access various data sources through the API interface to obtain data; The data cleaning unit is used to remove redundant and erroneous data by cleaning and preprocessing the collected data; The data storage unit is used to construct a data model of the knowledge graph by storing the mapped data in the knowledge graph; The human-computer interaction unit is used to display the knowledge graph through a user interface and receive user queries and instructions.

9. The multi-source heterogeneous big data fusion and reasoning system based on knowledge graph according to claim 7 is characterized in that: The workflow of the graph neural network module is as follows: First, obtain data from multiple data sources through web crawlers or API interfaces; Secondly, by defining the mapping rules between data sources and entities, the data from different data sources are mapped to a unified knowledge graph format; Then, the data embedding learning model learns the embedding vectors of nodes through training data and converts the knowledge graph of the graph structure into semantic vectors; Finally, the fused nodes are extracted through the embedding vector and mapping interface.

10. The multi-source heterogeneous big data fusion and reasoning system based on knowledge graph according to claim 7 is characterized in that: The workflow of the neural symbolic reasoning module is: First, the attribute information, text description information and knowledge of the node are represented as symbolic logic, that is, the attribute information and text description information of the fused node are converted into symbolic logic and feature vectors through the attention model; Secondly, the symbolic logic and the feature vector are input into a neural network based on symbolic logic to infer and generate new knowledge triples, that is, new knowledge triples are inferred based on the constraints contained in the symbolic logic and the information in the feature vector to achieve knowledge completion of the knowledge graph; Finally, the knowledge in the knowledge graph is updated according to the generated knowledge triples to achieve knowledge updating.

Images