Trusted knowledge graph node classification method and system based on instance classification difficulty

Abstract

The invention belongs to the field of knowledge graph completion, and discloses a trusted knowledge graph node classification method and system based on instance classification difficulty, and the method comprises the steps: obtaining a knowledge graph which uses a resource description framework standard, i.e., is organized in a triple form, and preprocessing the knowledge graph to obtain a graph structure data set of the knowledge graph; approximately estimating the classification difficulty of the nodes based on the graph structure attributes of the knowledge graph; under an ensemble learning framework, constructing a trusted knowledge graph node classification progressive learning model based on instance classification difficulty, and training the trusted knowledge graph node classification progressive learning model by using the nodes after classification difficulty approximate estimation to obtain an optimized trusted knowledge graph node classification progressive learning model; and inputting to-be-classified nodes and graph information into the optimized trusted knowledge graph node classification progressive learning model to obtain a classification result of knowledge graph node labels.

Description

Technical Field

[0001] The present invention belongs to the technical field of knowledge graph completion, and particularly relates to a trustworthy knowledge graph node classification method and system based on instance classification difficulty. Background Art

[0002] A knowledge graph is a structured knowledge representation form that organizes and stores the semantic relationships between various entities in the form of a graph, expressing the concepts in the physical world and their interrelationships.

[0003] In the past decade, numerous knowledge databases have been constructed, such as Freebase, Yago, Wikipedia, etc.; these knowledge databases play a key role in important fields of artificial intelligence such as semantic search and understanding, natural language processing, and knowledge representation. Through the application of knowledge graphs, artificial intelligence systems can better understand and reason about semantic information, achieving more accurate and intelligent search and understanding. Knowledge graphs also provide rich background knowledge for natural language processing, enabling computers to better understand and process natural language texts, while providing cross-domain knowledge fusion and application capabilities, enabling artificial intelligence systems to share and reuse knowledge in tasks in different fields.

[0004] However, due to the complexity of real-world knowledge and the diversity of actual business scenarios, challenges in terms of scale and quality are often faced during the construction and maintenance of knowledge graphs. Especially in large-scale knowledge graphs, there is a problem that a large amount of entity data lacks relevant category information, which leads to the incompleteness of the knowledge graph in certain fields or topics, thus restricting the utility and reliability of the knowledge graph in certain tasks and applications.

[0005] To address the incompleteness of knowledge graphs, the knowledge graph node classification task has become a popular research issue. Given a knowledge graph and a fixed label set, knowledge graph node classification needs to learn neighbor information to assign one or more category labels from the label set to the entities with missing labels in the graph.

[0006] However, there are two main problems in the current mainstream research on knowledge graph node classification, namely, the "black box" property of the representation learning method and the neglect of node difficulty information. First, these methods aggregate similar nodes and distance from different nodes according to neighbor information. However, these methods usually have the "black box" property, which destroys the interpretability and complex semantic structure of the knowledge graph itself, that is, it cannot directly provide an interpretable node feature representation, making the learned node representation difficult to explain and understand.

[0007] Secondly, many existing studies have shown that node classification performance is closely related to the difficulty information of the node. However, most existing knowledge graph node classification models treat all entities equally during the learning process, ignoring the impact of difficulty on classification performance. Specifically, "difficulty" generally refers to the degree to which it is difficult to distinguish instance categories when instances of different categories are mixed together; in most cases, instances with higher classification difficulty are more difficult to classify and contribute less to classification performance. This results in the difficulty of node classification potentially varying greatly depending on the semantic and structural environment in which they are located, that is, some nodes may be easier to correctly classify, while other nodes may be more challenging. Ignoring the difficulty of node classification may prevent the classifier from fully utilizing the intrinsic characteristics of each node in the knowledge graph, thereby affecting the improvement of classification performance.

[0008] In order to solve the above problems, the present invention proposes a new trusted knowledge graph node classification method named CNCKG to solve the problems existing in the existing technology. Summary of the Invention

[0009] In order to solve the above technical problems, the present invention provides a trusted knowledge graph node classification method and system based on instance classification difficulty, organizes and represents knowledge graphs in a structured manner, and proposes a novel, efficient, highly interpretable and cost-effective method for trusted completion of knowledge graph entity attribute information.

[0010] A trustworthy knowledge graph node classification method based on instance classification difficulty, the method comprising:

[0011] Obtain a knowledge graph organized in the form of triples using the Resource Description Framework standard, and preprocess the knowledge graph to obtain a graph structure dataset of the knowledge graph;

[0012] Approximately estimate the classification difficulty of nodes based on the graph structure attributes of the knowledge graph;

[0013] Under the ensemble learning framework, a trustworthy knowledge graph node classification progressive learning model based on instance classification difficulty is constructed. The nodes after approximate estimation of classification difficulty are used to train the trustworthy knowledge graph node classification progressive learning model, and the optimized trustworthy knowledge graph node classification progressive learning model is obtained.

[0014] The nodes and graph information to be classified are input into the optimized trusted knowledge graph node classification progressive learning model to obtain the classification results of the knowledge graph node labels.

[0015] Preferably, the classification difficulty of nodes is estimated approximately based on the graph structure attributes of the knowledge graph, including:

[0016] By analyzing the graph structure characteristics of the knowledge graph, information about nodes and their mutual relationships is obtained, and then it is used to calculate the approximate classification difficulty of nodes. Among them, Katz centrality is used as an index to measure the approximate difficulty of nodes.

[0017] Preferably, the trustworthy knowledge graph node classification progressive learning model includes: an interpretable node feature representation module OFR that retains the original semantic information and a knowledge graph node label classification module CPLD;

[0018] Among them, the interpretable node feature representation module OFR that retains the original semantic information adopts the method of retaining all the original information of the neighborhood with a specified depth of the node of interest, so that the learning process of the obtained nodes and relationships is kept interpretable with the feature representation; it also supports a low-cost and interpretable update strategy for retaining the original graph data to dynamically adjust the feature representation of nodes;

[0019] The knowledge graph node label classification module CPLD adopts an improved progressive ensemble learning algorithm based on instance classification difficulty; this algorithm is based on the Boosting learning framework, uses the method of base classifiers to learn the learned knowledge graph node feature representation, and dynamically adjusts the learning weights according to the classification difficulty of the misclassified instances.

[0020] Preferably, inputting the node to be classified and graph information into the optimized trustworthy knowledge graph node classification progressive learning model, the classification result of the knowledge graph node label obtained includes:

[0021] Obtaining the node feature representation of the node to be classified based on the interpretable node feature representation module OFR that retains the original semantic information;

[0022] Obtaining the classification difficulty of the node to be classified based on the low-cost difficulty calculation method;

[0023] Obtaining the completed knowledge graph node label based on the node feature representation, node classification difficulty, and the knowledge graph node label classification module CPLD.

[0024] The present invention also provides a trustworthy knowledge graph node classification system based on instance classification difficulty. The system is used to implement the foregoing method, including a preprocessing module, an estimation module, a training module, and a classification module;

[0025] The preprocessing module is used to obtain a knowledge graph organized in the form of triples using the Resource Description Framework standard, and preprocess the knowledge graph to obtain a graph structure data set of the knowledge graph;

[0026] The estimation module is used to approximately estimate the classification difficulty of nodes based on the graph structure attributes of the knowledge graph;

[0027] The training module is used to construct a progressive learning model for trustworthy knowledge graph node classification based on instance classification difficulty under the integrated learning framework, and use the nodes after approximate estimation of classification difficulty to train the progressive learning model for trustworthy knowledge graph node classification, so as to obtain an optimized progressive learning model for trustworthy knowledge graph node classification;

[0028] The classification module is used to input the node to be classified and graph information into the optimized progressive learning model for trustworthy knowledge graph node classification to obtain the classification result of the knowledge graph node label.

[0029] Preferably, the approximate estimation of the classification difficulty of nodes based on the graph structure attributes of the knowledge graph includes:

[0030] By analyzing the graph structure characteristics of the knowledge graph, information about nodes and their mutual relationships is obtained, and then used to calculate the approximate classification difficulty of nodes. Among them, Katz centrality is used as an index to measure the approximate difficulty of nodes.

[0031] Preferably, the progressive learning model for trustworthy knowledge graph node classification includes: an interpretable node feature representation module OFR that retains the original semantic information and a knowledge graph node label classification module CPLD;

[0032] Among them, the interpretable node feature representation module OFR that retains the original semantic information adopts the method of retaining all the original information of the neighborhood with a specified depth of the node of interest, so that the learning process of the obtained nodes and relationships is kept interpretable with the feature representation; it also supports a low-cost and interpretable update strategy for retaining the original graph data to dynamically adjust the feature representation of nodes;

[0033] The knowledge graph node label classification module CPLD adopts an improved progressive integrated learning algorithm based on instance classification difficulty; this algorithm is based on the Boosting learning framework, uses the way of base classifiers to learn the learned knowledge graph node feature representation, and dynamically adjusts the learning weights according to the classification difficulty of the misclassified instances.

[0034] Preferably, inputting the node to be classified and graph information into the optimized progressive learning model for trustworthy knowledge graph node classification to obtain the classification result of the knowledge graph node label includes:

[0035] Obtaining the node feature representation of the node to be classified based on the interpretable node feature representation module OFR that retains the original semantic information;

[0036] Obtaining the classification difficulty of the node to be classified based on the low-cost difficulty calculation method;

[0037] Obtaining the completed knowledge graph node label based on the node feature representation, node classification difficulty and the knowledge graph node label classification module CPLD.

[0038] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, the foregoing method is implemented.

[0039] The present invention also provides a computer-readable storage medium storing a computer program, which when executed, implements the foregoing method.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] The present invention not only focuses on the overall performance such as the error rate of the classifier, but also focuses on the credibility of each node prediction. Credibility is a metric for measuring the performance of a classifier from the perspective of the classification ability of a single node. According to the credibility theory, simple instances are more useful than difficult instances. It is unacceptable for a trustworthy classifier to misclassify low-difficulty nodes, while relatively difficult nodes are acceptable. Based on such a view, CNCKG focuses on the classification results of low-difficulty nodes and constructs a progressive learning model from easy to difficult with excellent accuracy and credibility under the framework of ensemble learning.

[0042] The present invention adopts an interpretable node feature representation method capable of retaining the original semantic information of the knowledge graph, and uses the technology of converting the neighbors of the nodes in the knowledge graph into a binary matrix to construct the embedding representation. These embedding representations can be combined with machine learning techniques to produce results that are both competitive and easy to interpret.

[0043] In order to reduce the computational cost of calculating the classification difficulty of nodes in the credibility evaluation index, the present invention proposes a low-cost difficulty calculation scheme from the perspective of the inherent structural attributes of the knowledge graph itself to approximately estimate the classification difficulty of nodes.

[0044] The present invention realizes a simple, efficient, interpretable and trustworthy method for complementing the entity attribute information of the knowledge graph, enriches and improves the current technical system for complementing the entity attribute information of the knowledge graph, helps to further explore the hidden knowledge of the knowledge graph, and improves the accuracy and precision of the application of the knowledge graph. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the technical solutions of the present invention, the following briefly introduces the drawings required in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0046] Figure 1It is the flowchart of the method in the embodiment of the present invention;

[0047] Figure 2 It is the schematic diagram of classifying the difficulty of approximately estimating nodes by the low-cost difficulty calculation method in the embodiment of the present invention;

[0048] Figure 3 It is the schematic diagram of an interpretable node feature representation that preserves the original semantic information in the embodiment of the present invention;

[0049] Figure 4 It is the schematic diagram of the overall conversion process of an interpretable node feature representation that preserves the original semantic information in the embodiment of the present invention;

[0050] Figure 5 It is the schematic diagram of the dynamic transformation process of an interpretable node feature representation that preserves the original semantic information in the embodiment of the present invention;

[0051] Figure 6 It is the overall structure diagram of the progressive learning model for knowledge graph entity information completion based on instance classification difficulty in the embodiment of the present invention;

[0052] Figure 7 It is the schematic diagram of the trend analysis of classification difficulty and approximate difficulty in the embodiment of the present invention. Detailed implementation manners

[0053] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0054] It should be noted that unless otherwise defined, the technical terms or scientific terms used in the embodiments of the present disclosure should have the ordinary meaning understood by those of ordinary skill in the art to which the present disclosure belongs. The "first", "second" and similar terms used in the embodiments of the present disclosure do not indicate any order, quantity or importance, but are only used to distinguish different components. The terms such as "including" or "comprising" mean that the elements or objects appearing before the term cover the elements or objects listed after the term and their equivalents, without excluding other elements or objects. The terms such as "connected" or "coupled" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. The terms such as "upper", "lower", "left" and "right" are only used to represent relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0055] First, some technical terms used in the present invention are elaborated:

[0056] The Resource Description Framework (RDF for short) is a markup language for describing network resources. It uses XML syntax to represent data models and is mainly used to describe the characteristics of Web resources and the relationships between them. The basic idea of RDF is that any network resource can be represented by a unique URI (Uniform Resource Identifier), which can be simplified to the URL (Uniform Resource Locator) here. In RDF, the resource is the subject, the property name is the predicate, and the property value is the object. This structure is called the "subject-predicate-object" structure or triple. The core advantage of RDF lies in its universality. It allows different vocabularies to be used to describe various types of resources, such as Web pages, books, online shopping products, etc. This enables information to be exchanged between applications while maintaining its semantic integrity, thereby improving the accuracy and intelligence of searches. In addition, RDF also defines RDF Schema (RDFS) for describing metadata property elements, such as "creator", which provides a basic vocabulary for describing resources. Generally speaking, RDF is a powerful tool for representing, exchanging, and processing metadata on the network, making the processing of information more efficient and intelligent.

[0057] An instance is an actual case or specific example used to prove the correctness or feasibility of a certain view, theory, law, etc. In daily life and academic discussions, instances are widely used to specifically elaborate and illustrate abstract concepts or principles. 1 Instances in daily language: For example: "The shortage of firewood is the most obvious and striking instance of the consequences of deforestation." 2 Instances in academic and scientific fields: In scientific research and theory verification, scientists often provide instances through experiments or observations to support or refute a certain hypothesis or theory. 3 Instances in object-oriented programming: In object-oriented programming (OOP), a class is an abstract data type, and an instance is the concretization of a class. Each instance is an object of the class and has the attributes and methods defined by the class. 4 Instances in databases: In database management, an instance usually refers to the set of programs that support the operation of a database. For example, in an Oracle database, an instance is a running instance of a database management system (DBMS), responsible for handling operations such as database creation, query, update, and deletion. 5 Instances in cloud services: In cloud computing, an instance refers to an independent computing unit allocated on a cloud server, and users can access and use various resources and services provided by the cloud service through the instance. Through these different application scenarios, we can see that instances play an important role in multiple fields, helping people better understand and apply abstract concepts.

[0058] An ensemble learning framework is a technical approach that improves overall performance by combining multiple base models. The core idea of ensemble learning is to utilize the advantages of multiple models and fuse them together through a certain strategy to obtain a more powerful and accurate prediction result than a single model. This method can reduce prediction errors, improve the robustness and resilience of the model, and better adapt to complex data distributions and uncertainties. Ensemble learning frameworks can be classified into the following types: 1. Bagging: Bagging, also known as bootstrap aggregating, is a parallel ensemble learning method. It randomly samples with replacement from the training set to generate multiple base models (individual learners), then trains these base models separately, and finally makes a comprehensive judgment based on the prediction results of all base models to obtain the final prediction result. 2. Boosting: Boosting is a sequential ensemble learning method that focuses on samples mispredicted by the previous model by gradually building models. Each subsequent model attempts to correct the errors of the previous model, and combines the outputs of these models through weighted voting or averaging the prediction results. 3. Stacking: Stacking is a meta-learning method that trains a meta-model to combine the prediction results of multiple base models. The meta-model is usually a learner that can learn from the predictions of the base models and make the final prediction. Ensemble learning frameworks have a wide range of applications and can be used for tasks such as classification, regression, and clustering, and have achieved significant performance improvements in many practical applications. By reasonably selecting and applying different ensemble learning frameworks, the performance and reliability of machine learning models can be effectively improved.

[0059] To make the above objects, features, and advantages of the present invention more obvious and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] Embodiment 1

[0061] As can be seen from the background art,

[0062] There are two main problems in the current mainstream research on knowledge graph node classification, namely, the "black box" property of representation learning methods and the neglect of node difficulty information. First, these methods aggregate similar nodes and distance different nodes based on neighbor information. However, these methods usually have the "black box" property, which destroys the interpretability and complex semantic structure of the knowledge graph itself, that is, they cannot directly provide interpretable node feature representations, making the learned node representations difficult to interpret and understand.

[0063] Secondly, many existing studies have shown that node classification performance is closely related to the difficulty information of the node. However, most existing knowledge graph node classification models treat all entities equally during the learning process, ignoring the impact of difficulty on classification performance. Specifically, "difficulty" generally refers to the degree to which it is difficult to distinguish instance categories when instances of different categories are mixed together; in most cases, instances with higher classification difficulty are more difficult to classify and contribute less to classification performance. This results in the difficulty of node classification potentially varying greatly depending on the semantic and structural environment in which they are located, that is, some nodes may be easier to correctly classify, while other nodes may be more challenging. Ignoring the difficulty of node classification may prevent the classifier from fully utilizing the intrinsic characteristics of each node in the knowledge graph, thereby affecting the improvement of classification performance.

[0064] like Figure 1 As shown, this embodiment provides a trusted knowledge graph node classification method, specifically providing a trusted knowledge graph node classification progressive learning model, including the following steps:

[0065] Obtain a knowledge graph organized in the form of triples (subject-verb-object) using the Resource Description Framework standard, and preprocess the knowledge graph to obtain a graph structure dataset of the knowledge graph;

[0066] The low-cost node classification difficulty calculation scheme proposed by this method uses Katz centrality to approximate the classification difficulty of the node;

[0067] Under the ensemble learning framework, a trusted knowledge graph node classification progressive learning model CNCKG based on instance classification difficulty is constructed. The trusted knowledge graph node classification progressive learning model includes an interpretable node feature representation module OFR that retains the original semantic information and a knowledge graph node label classification module CPLD. The node feature representation module and the knowledge graph node label classification module are used to guide the training of the progressive learning model, and the optimized trusted knowledge graph node classification progressive learning model is obtained.

[0068] The nodes and graph information to be classified are input into the trained trustworthy knowledge graph node classification progressive learning model to obtain the classification results of the knowledge graph node labels.

[0069] The specific implementation process includes:

[0070] Step 1: Obtain the knowledge graph using the Resource Description Framework standard, that is, organize it in the form of triples (subject-verb-object), preprocess the knowledge graph, and obtain the knowledge graph structure and all label sets of the entities;

[0071] Step 2: Figure 2As shown in the figure, a low-cost node classification difficulty calculation scheme is adopted to approximately estimate the classification difficulty of nodes and add it as prior knowledge to node classification; the Katz centrality is used as an indicator to measure the approximate difficulty of nodes in the scheme. The Katz centrality not only reflects the structural characteristics and aggregation characteristics of nodes in the graph, but also comprehensively considers the connection patterns and path lengths between nodes and their direct and indirect neighbors. This reflects the local class overlap of nodes in the graph, thus partially revealing the aggregation properties of nodes and their propagation ability in the graph, while significantly improving the calculation efficiency. The Katz centrality of a node is defined as:

[0072]

[0073] where c k (v i ) represents the approximate difficulty of node v i , A represents the adjacency matrix of the knowledge graph, and θ and η are difficulty parameters.

[0074] Step 3: Obtain the node feature representation of the node to be classified through the interpretable node feature representation module OFR that preserves the original semantic information; as Figure 3 shown in the figure, to obtain the dictionary feature representation result at a given depth d, the representation learning method first queries the neighborhood information of all interested nodes, that is, the nodes to be classified, at a query depth of 1, and stores this neighborhood information using a dictionary. Then, for each given interested node, the complete neighborhood dictionary is constructed in an iterative manner. When the node depth is 1, all the edges in the neighborhood of the interested node are used as keys, and their corresponding target nodes or constants are used as values to form multiple key-value pairs, which are inserted into the neighborhood dictionary of the interested node; next, query the neighborhood nodes at a depth of 2 for the interested node, and connect the edge paths between the original node and these newly found nodes as keys, and use these newly found nodes as values to form new key-value pairs, and add them to the neighborhood dictionary of the original node. Repeat this process iteratively until the depth d is reached. In order to clearly represent in the neighborhood dictionary the neighborhood of the interested node that contains certain edges, the representation learning method will sequentially concatenate all the edges on the paths from the interested node to other target attributes or nodes at the current depth. Then, these edges and "Exist" are used as values to form multiple key-value pairs, which are inserted into the neighborhood dictionary. Thus, the result as Figure 3The neighborhood dictionary of all interesting nodes. Assume that h1 is an interesting node, and the given depth d = 2. In the first iteration l = 1, the learning method extracts the nodes h2, h3, h5, h6 with a one-hop neighborhood, and then forms key-value pairs r1: h2, r2: h3, r3: h5, r5: h6 with their edge necklaces and adds them to the neighborhood dictionary of h1. In the second iteration l = 2, the learning method extracts the node h4 with a two-hop neighborhood, and then connects the edges from the original node to these nodes with dots r1.r4, and forms a key-value pair r1.r4: h4 with h4 and adds it to the neighborhood dictionary of h1. The final result is as Figure 4 shown. Each interesting point forms its own neighborhood dictionary at the given depth. The dictionaries are transformed to form a representation learning matrix, which contains the fully interpretable structure and attribute information of the knowledge graph and maintains the ability to expand the existing dataset, and will be used for subsequent node classification methods.

[0075] Step 4: Through the dictionary representation of the node neighborhood, use a dynamic update strategy to incrementally and dynamically adjust the feature representation of the nodes. The dynamic update strategy aims to control the impact of node / relationship updates in the knowledge graph locally rather than globally, that is, without recalculating the existing feature representations, efficiently adapt to the changes in the graph structure, and maintain the accuracy, consistency, and interpretability of the feature representations. Specifically, when the graph data of the knowledge graph increases, deletes, or modifies, etc., the scope that needs to be adjusted is limited to the feature representations of neighboring nodes. As Figure 5 shown, when new nodes and relationships are added to the graph data, the original graph data has obtained an embedding matrix through representation learning, and all key-value pairs in the form of are recorded as a set in the matrix. The dynamic update strategy will first calculate the set of key-value pairs generated by adding nodes and relationships. Then calculate the feature columns that need to be added in the embedding matrix. Search for neighboring nodes within the neighborhood (set depth d) of the newly added node, supplement the feature vector of the newly added node according to the interpretable feature representation method, and at the same time supplement the values of these neighboring nodes under the new feature, while the values of non-neighboring nodes under the new feature are defaulted to 0. This strategy effectively avoids retraining the entire knowledge graph. Similarly, if a deletion operation occurs in the graph data, it is necessary to calculate to determine the feature columns that need to be deleted in the embedding matrix. This dictionary learning-based method not only significantly reduces the computational cost and time consumption compared with retraining the entire embedding model when dealing with graph data updates, but also does not destroy the interpretability of the representation learning due to the update of the graph data because it maintains the accuracy and consistency of the embedding.

[0076] Step 5: The knowledge graph node label classification module proposes to adopt the improved progressive learning algorithm CPLD based on instance classification difficulty; this algorithm is based on the Boosting learning framework, trains the base classifier based on the learned knowledge graph node feature representation, and dynamically adjusts the learning weights according to the classification difficulty of the misclassified nodes. Without reducing the classification accuracy, it reduces the error rate of difficult nodes and reduces their impact on the generalization performance of the classifier. This algorithm starts training from the nodes with the lowest difficulty, trains multiple base learners in sequence from easy to difficult for the training nodes, and then uses the weighted voting method to aggregate the classification results of the classifier. The misclassified nodes with lower classification difficulty will increase their weights in the next iteration and are more likely to be selected. This algorithm distinguishes instances based on instance difficulty and reduces the risk of overfitting.

[0077] The formula for calculating the misclassification cost in this algorithm is:

[0078]

[0079] where ∈ represents the misclassification parameter, and the larger the value of ∈, the closer the value of i is to 0. In addition, if the classification difficulty value d of an instance is equal to 0, then the misclassification loss i of this instance reaches the maximum value, which is 1. Therefore, the value range of d is (0, 1]. Since CPLD is a voting-based ensemble learning algorithm, assuming the classifier set is defined as

[0080]

[0081] H(v i ) = argmax k f k (v i ) (4)

[0082] where α t and β t (v i ) represent the learning weight and the prediction result of the base classifier β t respectively, f k (v i ) represents the weighted voting score of node v i for class k, and H(v i ) is the predicted label of node v i . During the iteration process, the CPLD algorithm hopes to focus more on correctly classifying instances with lower classification difficulty in the next iteration. Therefore, its weight update rule is:

[0083]

[0084] where ω i，t+1 is the weight of node v i in the t-th iteration, y i is the correct label of node v i , and the normalization factor z t is expressed as:

[0085]

[0086] From the above formula, it can be known that if an instance v i with a lower classification difficulty value d i is misclassified, its misclassification cost is relatively large. Therefore, in the next iteration, the weight ω i，t+1 will increase, and the probability of its being sampled will also increase.

[0087] In each iteration, the CPLD algorithm must ensure the weighted distribution of the dataset. To this end, the initial weighted distribution of the dataset v is defined as follows:

[0088]

[0089] The CPLD algorithm aims to reduce the training loss of instances without compromising classification accuracy. Therefore, the objective function of the CPLD algorithm is as follows:

[0090]

[0091] Through the above five steps, the original information of the knowledge graph and machine learning techniques can be used to perform trustworthy classification on nodes of unknown categories, as Figure 6 shown.

[0092] Example 2

[0093] This example provides a verification test for trustworthy completion of entity attribute information in a knowledge graph, including:

[0094] Step 1: Select five commonly used knowledge graph datasets for difficulty calculation and performance analysis of methods. These knowledge graphs are RDF-formatted datasets collected by Ristoski, which have been pre-divided into training sets and test sets. They are: BGS, AIFB, MUTAG, AM, and DBpedia Cities. Among them, BGS has 146 binary-labeled rock entities, described by more than 100k triples, with 105 relationship types on the graph, and each node has an average of 280 predicate-object pairs. AIFB has 178 research group entities for a three-classification problem, approximately 29k triples, and 47 different relationship types. MUTAG describes the information of composite molecules that may be carcinogenic, with the label of mutagenicity (129 positive samples, 211 negative samples), a binary classification problem. The dataset has 75K triples and 24 unique relationship types. AM describes the information of artifacts in the Amsterdam Museum, with a total of 1000 samples for 11 categories of classification. The dataset has 57 million triples and 100 different relationships. Finally, DBpedia Cities describes the information of 212 urban living indicators collected before 2012, and is divided into three categories: high, medium, and low, with 67, 106, and 39 entities respectively.

[0095] Step 2: Experimental settings. It is implemented in the Pytorch 1.10.0 framework on a Windows 11 machine and trained and tested on a platform equipped with an RTX 3090 (24GB) and a 14vCPU Intel(R) Xeon(R) Gold 6330 CPU @ 2.00GH. Five types of experiments are conducted on this platform: verification of the effectiveness of the node approximation difficulty scheme, performance evaluation of the node classification method and the comparison method in terms of accuracy and Cr, comparison of the running times of different node classification methods, and verification of the acceptability of classifiers and the effectiveness of dynamic updates on different knowledge graph datasets. When verifying the effectiveness of the approximation difficulty scheme, 20 classifiers are selected to calculate the statistical classification difficulty of nodes in the five datasets, and a tenth-degree polynomial fitting is performed with the approximate difficulty calculated by the approximate classification difficulty scheme. Each algorithm in all experiments is repeated five times with ten-fold cross-validation, and the average value is taken as the result.

[0096] Step 3: Evaluate the experimental results.

[0097] The experiment on verifying the effectiveness of node approximation difficulty fits the statistical classification difficulty and the approximate classification difficulty on the five datasets, from Figure 7It can be intuitively seen that the distribution trends of the approximate difficulty of the vast majority of nodes in each dataset are very similar to the distribution trends of the statistical classification difficulty. The above results indicate that it is feasible to design a reasonable approximate difficulty substitution method based on the graph structure and semantic information characteristics of the knowledge graph, and it can express the information of node classification difficulty to a certain extent. At the same time, since a large number of classifiers are no longer needed to count the classification results, the calculation efficiency of the difficulty value is greatly improved.

[0098] Six comparison methods were used to quantitatively evaluate the error rate of the proposed trustworthy knowledge graph entity classification model CNCKG in each difficulty interval, as well as the accuracy and Cr values of each dataset, for the performance evaluation of the node classification method and the comparison method in terms of accuracy and Cr. The comparison methods include the INK (Instance Neighbouring by using Knowlede) algorithm, the RDF2Vec algorithm, the AdaBoost (Adaptive Boosting) algorithm, the RGCN (Relational Graph Convolutional Networks) algorithm, the ERGCN (Data Enhancement-based Robust Graph Convolutional Network) algorithm, and the CGPN (Multi-relational Graph Gaussian Process network) algorithm. The average misclassification rates of CNCKG and the six types of comparison methods in different difficulty intervals on five datasets are shown in Table 1. Generally speaking, as the node classification difficulty increases, the misclassification rate of the CNCKG algorithm shows a relatively stable growth trend. For other models, this trend is either weak or non-existent. In the low-difficulty interval of [0, 0.6], the misclassification rate of the CNCKG algorithm is much lower than that of the other six comparison methods. In the high-difficulty interval of (0.6, 1.0], the error rate of the CNCKG algorithm is higher than that of the other six comparison algorithms, and the misclassification rate of CNCKG in the low-difficulty interval of [0, 0.6] is significantly lower than that in the high-difficulty interval of (0.6, 1.0]. The CNCKG algorithm basically conforms to the natural cognitive principle: the misclassification rate of simple nodes is lower than that of difficult nodes, and it effectively reduces the attention to difficult nodes, thereby slowing down the overfitting problem to a certain extent, as shown in Table 1.

[0099] Table 1 Average error rates of comparison methods on five datasets in different difficulty intervals

[0100]

[0101] The average accuracy and confidence results of six comparison algorithms on five datasets are shown in Tables 2 and 3 respectively. The experimental results marked with an asterisk (*) indicate that our experimental equipment could not obtain valid results, and these results are cited from the original literature. The results in Tables 2 and 3 show that CNCKG achieved the highest classification accuracy on four datasets and the highest confidence on five datasets. Specifically, CNCKG achieved the best accuracy results on the Mutag, AM, BGS, and Movies datasets, with improvements of 1.5%, 3.0%, 1.9%, and 1.2% respectively compared to the second-best results. In the AIFB dataset, CNCKG ranked second, second only to the best-performing RGCN, lagging behind by 0.9%. In terms of confidence, CNCKG achieved the highest results on the AIFB, Mutag, AM, BGS, and Movies datasets, with improvements of 0.6%, 1.6%, 1.7%, 3.7%, and 2.5% respectively compared to the second-best results. In addition, when comparing CNCKG with the ablation method Adaboost, significant advantages in both accuracy and confidence were shown on all five datasets. This indicates that the node difficulty information derived from the confidence theory does contribute to improving the performance and confidence of knowledge graph node classification.

[0102] Table 2 Accuracy results on five datasets (± represents variance, the best result is in bold, and the second-best is underlined)

[0103] Model AIFB Mutag AM BGS Movies INK 88.9(±0.00) 77.9(±0.00) 86.8(±0.00) 92.1(±0.01) 91.1(±0.01) RDF2Vec 91.7(±0.00) 80.5(±0.01) 88.2(±0.01) 83.4(±0.02) 73.2(±0.02) AdaBoost 89.2(±0.01) <![CDATA 81.8(±0.01) > 87.3(±0.01) <![CDATA 92.6(±0.01) > 91.9(±0.01) RGCN 95.3(±0.01) 69.1(±0.02) <![CDATA 88.7(±0.01) > 78.5(±0.02) <![CDATA 93.9(±0.01) > ERGCN 91.6(±0.06) 74.2(±0.01) 88.1(±0.01) 84.4(±0.03) 93.3(±0.01) GGPN 92.1(±0.01) 77.9(±0.03) 87.4* 85.0(±0.01) 62.1(±0.01) CNCKG <![CDATA 94.4(±0.01) > 83.3(±0.01) 91.7(±0.01) 94.5(±0.01) 95.1(±0.01) impr. -0.90％ 1.50％ 3.00％ 1.90％ 1.20％

[0104] Table 3 Confidence results on five datasets (± represents variance, the best result is in bold, and the second-best is underlined)

[0105] Model AIFB Mutag AM BGS Movies INK 88.5(±0.00) 90.8(±0.00) 89.2(±0.05) 92.4(±0.08) 91.8(±0.08) RDF2Vec 94.3(±0.00) 90.4(±0.02) 91.3(±0.01) 90.3(±0.01) 89.9(±0.00) AdaBoost 93.9(±0.01) <![CDATA 94.9(±0.01) > <![CDATA 96.0(±0.01) > <![CDATA 95.2(±0.03) > 95.1(±0.01) RGCN <![CDATA 97.7(±0.01) > 83.9(±0.01) 92.2(±0.01) 87.3(±0.01) 93.3(±0.01) ERGCN 94.5(±0.05) 86.0(±0.05) 91.4(±0.01) 96.3(±0.01) <![CDATA 95.9(±0.02) > GGPN 94.6(±0.01) 90.8(±0.02) * 95.1(±0.02) 78.4(±0.01) CNCKG 98.3(±0.01) 96.5(±0.01) 97.7(±0.01) 98.9(±0.01) 98.4(±0.01) impr. +0.6％ +1.6％ +1.7％ +3.7％ +2.5％

[0106] The running time comparison of different node classification methods was benchmarked by conducting five rounds of tests on each of the five datasets to compare the running time of CNCKG with the baseline models. The results are shown in Table 4. CNCKG showed the fastest running time on the BGS dataset and competitive performance on the Mutag and AM datasets. Although the running times on AIFB and Movies were longer, they were still competitive compared to other classifiers. As an interpretable knowledge graph node classification method, investing additional time in the upstream task to preserve the interpretability results of the original graph data contributes to its scalability in various practical applications. Therefore, considering the superior performance of CNCKG and Cr, the trade-off in running time is reasonable and acceptable for practical applications. The computational time overhead of the CNCKG algorithm is mainly affected by the dataset size and the depth used in the node representation based on the original information. In terms of dataset size, the Movies and AM datasets are the largest, resulting in higher time consumption for CNCKG. Regarding the depth, AIFB and Movies have a greater depth compared to Mutag, AM, and BGS, leading to relatively more time consumption. On the contrary, AM and BGS utilize the lowest depth, and thus, despite their larger dataset sizes, they require less computational time, as shown in Table 4.

[0107] Table 4 Running times of the comparison algorithms on five datasets

[0108] Model AIFB Mutag AM BGS Movies INK 6432.7 160.9 653.3 377214.5 27089.9 RDF2Vec 1593.7 2049.6 9361.4 538.0 1307.8 AdaBoost 4986.3 188.75 644.1 384376.3 26827.2 RGCN 20.1 383.0 8873.3 3329.3 37.4 ERGCN 136.1 73.7 7751.7 260.0 79.3 GGPN 88.4 78.9 * 45442.8 14930.9 CNCKG 4836.5 280.75 706.87 30.275 26587.7 INK 6432.7 160.9 653.3 377214.5 27089.9

[0109] After selecting a classifier, if there is no need to try other classifiers, then that classifier is considered a credible classifier. According to the theory of credible classifiers, if the confidence value of the trained classifier is not lower than the threshold θ, these classifiers are considered acceptable classifiers and can be used as substitutes for credible classifiers. The credibility value of each classifier is denoted as Cr. Let be the threshold, where is the average difficulty of all nodes in the knowledge graph. For any classifier, if Cr > θ, then the classifier is considered an acceptable classifier for the knowledge graph. The acceptable classifiers on different knowledge graph datasets are obtained by calculating the average difficulty of five knowledge graph datasets to obtain the acceptable threshold for each dataset. Then, based on the Cr results of CNCKG and the other six baselines, we can easily obtain the acceptable classifiers for each dataset (application scenario). For the AIFB dataset, the acceptable classifiers are RGCN and CNCKG; for the Mutag dataset, the acceptable classifiers include AdaBoost and CNCKG; for the BGS dataset, the acceptable classifiers are INK, AdaBoost, ERGCN, GGPN, and CNCKG; for the AM dataset, the acceptable classifiers include RDF2Vec, AdaBoost, RGCN, ERGCN, and CNCKG; for the Movies dataset, the acceptable classifiers are AdaBoost, ERGCN, and CNCKG. CNCKG is an acceptable classifier in all five datasets, which confirms its applicability in different scenarios.

[0110] Validation experiments on dynamic update effectiveness First, obtain the node representations based on the original information of the AIFB dataset and the MUTAG dataset as the baseline representations. Next, randomly delete ten nodes and their corresponding triples from each dataset to simulate the dynamic update operation. Then, obtain the node representations of the pruned KG, which is regarded as the KG before the update. Subsequently, add one node and its corresponding triple at a time, and apply the dynamic update strategy to obtain the latest node representations. Denote the original KG as OKG and the KG after incremental update as UKG. The results are shown in Table 5, indicating that at different depths, the diagonal cosine similarity between the node representations of OKG and UKG is 100%, which shows that the node representations obtained by the incremental update strategy are the same as those of the original KG, avoiding the recomputation of the feature representations of the entire KG. In addition, the prediction accuracy using the random forest classifier is almost the same for OKG and UKG, further verifying the effectiveness of the features obtained by dynamic update. When the depth is set to 1, the time to compute the entire representation of OKG is 5.45 seconds for AIFB and 25.49 seconds for MUTAG, while the average incremental update time for each node in UKG is 0.78 seconds for AIFB and 2.43 seconds for MUTAG. This shows that the incremental update significantly reduces the computational cost compared to recomputing the entire KG. The reduction in computational cost is due to fewer affected nodes at depth 1, so fewer new node features need to be computed. However, as the depth increases, for example, at depth 2, the computation time of OKG is 108.21 seconds for AIFB and 172.71 seconds for MUTAG, while the average update time for each node in UKG is 94.43 seconds for AIFB and 143.32 seconds for MUTAG. The reason is that at depths greater than 1, the AIFB and MUTAG KGs exhibit highly interconnected characteristics, resulting in the need to recompute the features of all nodes, thus reducing the magnitude of computational savings. Nevertheless, since the dynamic update strategy can insert nodes and directly compute new features without reloading the entire graph into memory, the computation time of UKG at depth 2 is still lower than that of OKG. It can be observed that at smaller depths, the magnitude of computational time savings is greater, while as the depth increases, the savings are relatively smaller. However, the dynamic update strategy ensures that as the number of updated nodes increases, the difference in computational time between OKG and UKG becomes larger. In addition, in most real-world scenarios, the update of a KG usually only affects a limited range of nodes, which means that not all nodes need to have their feature representations recomputed. Our method is based on this observation and aims to significantly reduce the cost of updating feature representations while ensuring their interpretability and effectiveness. Therefore, the dynamic update strategy also maintains the interpretability of CNCKG. Thus, the dynamic update strategy is expected to perform better in terms of efficiency in practical situations, as shown in Table 5.

[0111] Table 5 Comparison of OKG and UKG with different depths in terms of time, prediction accuracy, and diagonal cosine similarity

[0112]

[0113] As mentioned above, it is only the preferred specific implementation manner of the present application. However, the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0114] Example 3

[0115] This example provides an example of complementing the specific attributes of a device (node) in an industrial domain knowledge base in a device knowledge graph, including the following steps:

[0116] Step 1: Select the device knowledge graph that needs to be complemented with trustworthy entity attribute information. Industrial devices often have the following problems: frequent and multiple failures, the need to converge and precipitate maintenance knowledge and experience; industrial production processes and power dispatching processes involve multiple types of devices, and device fault analysis consumes a large amount of manpower and time; products / devices may have potential defects and need to be inspected for defects; industry and power have high requirements for the rigor and safety of products / devices, and relevant personnel are required to strictly check the devices involved in the overall industrial / power production process. Therefore, it is necessary to build a knowledge graph network based on past fault causes and knowledge manuals, and quickly locate the causes and measures according to fault phenomena. However, it is necessary to classify the importance and danger levels of devices. When constructing the knowledge graph, these two attributes usually need to be individually labeled by domain experts. In a knowledge graph with a large number of devices, these two important attributes often face the situation of attribute loss, and a trustworthy method for complementing knowledge graph entity attribute information is needed to improve the graph. After selecting the device knowledge graph, preprocess the knowledge graph to obtain the knowledge graph graph structure and the label sets of the importance and danger levels of the devices;

[0117] Step 2: Adopt a low-cost node classification difficulty calculation scheme to approximately estimate the classification difficulty of device nodes and add it as prior knowledge to the trustworthy complementation of the importance and danger level attributes of the devices; in the scheme, Katz centrality is used as an indicator to measure the approximate difficulty of device nodes.

[0118] Step 3: Use the interpretable node feature representation module OFR that retains the original semantic information to obtain the feature representation of each device node for the complementation of the importance and danger level attributes.

[0119] Step 4: Through the dictionary representation of the node neighborhood, use a dynamic update strategy to incrementally and dynamically adjust the feature representation of the nodes. If a new device is added to the knowledge graph or an existing device is deleted from the knowledge graph, there is no need to retrain the entire knowledge graph. Only local modifications are required.

[0120] Step 5: Use the improved progressive learning algorithm CPLD based on instance classification difficulty; this algorithm is based on the Boosting learning framework, trains the base classifier based on the learned knowledge graph node feature representation, and dynamically adjusts the learning weights according to the classification difficulty of the misclassified nodes.

[0121] Step 6: Input the feature representation of the device nodes lacking importance and danger levels into CPLD, and the missing importance and danger level attribute information can be obtained.

[0122] Through the above six steps, the trustworthy attribute information of the device nodes lacking importance and danger levels can be complemented using the original information of the knowledge graph and machine learning techniques. The expected goal can be achieved, that is, quickly discovering the causes of industrial equipment failures and taking corresponding measures according to the importance and danger levels to ensure the safety of industrial equipment and restore the normal use of the equipment.

[0123] Embodiment 4

[0124] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present disclosure also provides a trustworthy knowledge graph node classification system based on instance classification difficulty, including: a preprocessing module, an estimation module, a training module, and a classification module;

[0125] The preprocessing module is used to obtain a knowledge graph organized in the form of triples using the Resource Description Framework standard, and preprocess the knowledge graph to obtain a graph structure data set of the knowledge graph;

[0126] The estimation module is used to approximately estimate the classification difficulty of nodes based on the graph structure attributes of the knowledge graph;

[0127] The training module is used to construct a trustworthy knowledge graph node classification progressive learning model based on instance classification difficulty under the ensemble learning framework, and use the nodes approximately estimated by the classification difficulty to train the trustworthy knowledge graph node classification progressive learning model to obtain an optimized trustworthy knowledge graph node classification progressive learning model;

[0128] The classification module is used to input the node to be classified and the graph information into the optimized trustworthy knowledge graph node classification progressive learning model to obtain the classification result of the knowledge graph node label.

[0129] In this embodiment, the approximate estimation of the classification difficulty of nodes based on the graph structure attributes of the knowledge graph includes:

[0130] By analyzing the graph structure characteristics of the knowledge graph, information about nodes and the relationships between nodes is obtained, and then used to calculate the approximate classification difficulty of nodes. Among them, Katz centrality is used as an indicator to measure the approximate difficulty of nodes.

[0131] In this embodiment, the trustworthy knowledge graph node classification progressive learning model includes: an interpretable node feature representation module OFR that retains the original semantic information and a knowledge graph node label classification module CPLD;

[0132] Among them, the interpretable node feature representation module OFR that retains the original semantic information adopts retaining all the original information of the specified depth of the neighborhood of the nodes of interest, so that the learning process of the obtained nodes and relationships is kept interpretable with the feature representation; it also supports a low-cost and interpretable update strategy for retaining the original graph data to dynamically adjust the feature representation of nodes;

[0133] The knowledge graph node label classification module CPLD adopts an improved progressive ensemble learning algorithm based on the classification difficulty of instances; this algorithm is based on the Boosting learning framework, uses the way of base classifiers to learn the learned knowledge graph node feature representation, and dynamically adjusts the learning weights according to the classification difficulty of the misclassified instances.

[0134] In this embodiment, inputting the node to be classified and the graph information into the optimized trustworthy knowledge graph node classification progressive learning model, the classification result of the knowledge graph node label obtained includes:

[0135] Obtaining the node feature representation of the node to be classified based on the interpretable node feature representation module OFR that retains the original semantic information;

[0136] Obtaining the classification difficulty of the node to be classified based on the low-cost difficulty calculation method;

[0137] Obtaining the completed knowledge graph node label based on the node feature representation, the node classification difficulty, and the knowledge graph node label classification module CPLD.

[0138] The system of the above embodiment is used to implement the corresponding trustworthy knowledge graph node classification method based on the classification difficulty of instances in any one of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be elaborated here.

[0139] It should be noted that the above trustworthy knowledge graph node classification system based on the classification difficulty of instances is embodied in the form of functional units. The term "module" here can be implemented in software and / or hardware forms, and no specific limitation is made thereto.

[0140] For example, a "module" may be a software program, a hardware circuit, or a combination of both that implements the above functions. The hardware circuit may include an application specific integrated circuit (ASIC), an electronic circuit, a processor (such as a shared processor, a dedicated processor, or a group of processors, etc.) for executing one or more software or firmware programs, and a memory, a merged logic circuit, and / or other suitable components that support the described functions.

[0141] Embodiment Five

[0142] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present disclosure also provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements a method for classifying trustworthy knowledge graph nodes based on instance classification difficulty described in any one of the above embodiments.

[0143] The system of the above embodiments is used to implement a method for classifying trustworthy knowledge graph nodes based on instance classification difficulty corresponding to any one of the foregoing embodiments, and has the beneficial effects of the corresponding method embodiments, which will not be elaborated herein.

[0144] Embodiment Six

[0145] Based on the same inventive concept, corresponding to the method of any of the above embodiments, the present disclosure also provides a non-transitory computer-readable storage medium. The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to cause the computer to execute a method for classifying trustworthy knowledge graph nodes based on instance classification difficulty described in any one of the foregoing embodiments.

[0146] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. The information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory, or other memory technologies, compact disc read-only memory (CD-ROM), digital versatile disc (DVD), or other optical storage, magnetic cassette tapes, magnetic disk storage, or other magnetic storage devices, or any other non-transmission medium that can be used to store information accessible by a computing device.

[0147] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute a trustworthy knowledge graph node classification method based on instance classification difficulty as described in any one of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be elaborated here.

[0148] Embodiments of the present disclosure are intended to cover all such substitutions, modifications, and variations that fall within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present disclosure shall be included within the protection scope of the present disclosure.

Claims

1. A trustworthy knowledge graph node classification method based on the classification difficulty of instances, characterized in that The method includes: Obtaining a knowledge graph organized in the form of triples using the Resource Description Framework standard, and preprocessing the knowledge graph to obtain a graph structure dataset of the knowledge graph; Approximately estimating the classification difficulty of nodes based on the graph structure attributes of the knowledge graph; Under the framework of ensemble learning, constructing a progressive learning model for trustworthy knowledge graph node classification based on instance classification difficulty, and training the progressive learning model for trustworthy knowledge graph node classification with the nodes whose classification difficulty is approximately estimated to obtain an optimized progressive learning model for trustworthy knowledge graph node classification; Inputting the nodes to be classified and graph information into the optimized progressive learning model for trustworthy knowledge graph node classification to obtain the classification result of the knowledge graph node labels.

2. The method according to claim 1, characterized in that, Approximately estimating the classification difficulty of nodes based on the graph structure attributes of the knowledge graph includes: By analyzing the graph structure characteristics of the knowledge graph, obtaining information about nodes and the relationships between nodes, and then using it to calculate the approximate classification difficulty of nodes. Among them, Katz centrality is used as an indicator to measure the approximate difficulty of nodes.

3. The method according to claim 1, characterized in that, The progressive learning model for trustworthy knowledge graph node classification includes: an interpretable node feature representation module OFR that preserves the original semantic information and a knowledge graph node label classification module CPLD; Among them, the interpretable node feature representation module OFR that preserves the original semantic information adopts the method of preserving all the original information of the neighborhood with a specified depth of the nodes of interest, so that the learning process of the obtained nodes and relationships is kept interpretable to a certain extent; it also supports a low-cost and interpretable update strategy for preserving the original graph data to dynamically adjust the feature representation of nodes; The knowledge graph node label classification module CPLD adopts an improved progressive ensemble learning algorithm based on instance classification difficulty; this algorithm is based on the Boosting learning framework, uses the way of base classifiers to learn the feature representation of the knowledge graph nodes learned, and dynamically adjusts the learning weights according to the classification difficulty of the misclassified instances.

4. The method according to claim 3, wherein Inputting the nodes to be classified and graph information into the optimized progressive learning model for trustworthy knowledge graph node classification to obtain the classification result of the knowledge graph node labels includes: Obtaining the node feature representation of the nodes to be classified based on the interpretable node feature representation module OFR that preserves the original semantic information; Obtaining the classification difficulty of the nodes to be classified based on the low-cost difficulty calculation method; Obtaining the completed knowledge graph node labels based on the node feature representation, the node classification difficulty, and the knowledge graph node label classification module CPLD.

5. A trustworthy knowledge graph node classification system based on the classification difficulty of instances, the system is used to implement the method described in any one of claims 1-4, characterized in that, A preprocessing module, an estimation module, a training module, and a classification module; The preprocessing module is used to obtain a knowledge graph organized in the form of triples using the Resource Description Framework standard, and preprocess the knowledge graph to obtain a graph structure dataset of the knowledge graph; The estimation module is used to approximately estimate the classification difficulty of nodes based on the graph structure attributes of the knowledge graph; The training module is used to construct a progressive learning model for trustworthy knowledge graph node classification based on instance classification difficulty under the integrated learning framework, and use the nodes after approximate estimation of classification difficulty to train the progressive learning model for trustworthy knowledge graph node classification, so as to obtain an optimized progressive learning model for trustworthy knowledge graph node classification; The classification module is used to input the node to be classified and graph information into the optimized progressive learning model for trustworthy knowledge graph node classification, and obtain the classification result of the knowledge graph node label.

6. The system according to claim 5, wherein Approximately estimating the classification difficulty of nodes based on the graph structure attributes of the knowledge graph includes: By analyzing the graph structure characteristics of the knowledge graph, information about nodes and their mutual relationships is obtained, which is then used to calculate the approximate classification difficulty of nodes. Among them, Katz centrality is used as an index to measure the approximate difficulty of nodes.

7. The system according to claim 5, characterized in that, The progressive learning model for trustworthy knowledge graph node classification includes: an interpretable node feature representation module OFR that retains the original semantic information and a knowledge graph node label classification module CPLD; Among them, the interpretable node feature representation module OFR that retains the original semantic information adopts the method of retaining all the original information of the neighborhood with a specified depth of the node of interest, so that the learning process of the obtained nodes and relationships is kept somewhat interpretable with the feature representation; it also supports a low-cost and interpretable update strategy for retaining the original graph data to dynamically adjust the feature representation of nodes; The knowledge graph node label classification module CPLD adopts an improved progressive integrated learning algorithm based on instance classification difficulty; this algorithm is based on the Boosting learning framework, uses the way of base classifiers to learn the learned knowledge graph node feature representation, and dynamically adjusts the learning weights according to the classification difficulty of the misclassified instances.

8. The system according to claim 7, wherein, Inputting the node to be classified and graph information into the optimized progressive learning model for trustworthy knowledge graph node classification and obtaining the classification result of the knowledge graph node label includes: Obtaining the node feature representation of the node to be classified based on the interpretable node feature representation module OFR that retains the original semantic information; Obtaining the classification difficulty of the node to be classified based on the low-cost difficulty calculation method; Obtaining the completed knowledge graph node label based on the node feature representation, node classification difficulty, and the knowledge graph node label classification module CPLD.

9. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the program, it implements the method according to any one of claims 1 to 4.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and when the computer program is executed, it implements the method according to any one of claims 1 to 4.

Images