A trusted knowledge graph entity classification method based on a multi-view integrated model

Abstract

The application provides a trusted knowledge graph entity classification method based on a multi-view integrated model, which comprises the following steps: acquiring an adjacency matrix and entity feature data in a knowledge graph network, and generating an original view of knowledge graph data; calculating a knowledge graph entity adjacency matrix according to knowledge graph correlation information, and calculating a knowledge graph entity feature matrix according to a knowledge graph entity attribute relationship; constructing a structure view for the knowledge graph adjacency matrix, and constructing a feature view for the knowledge graph entity feature matrix; inputting the structure view and the original view into a structure-based model to output a first feature; inputting the feature view and the original view into a feature-based model to output a second feature; and inputting the first feature and the second feature into a linear layer to integrate and output the category of the knowledge graph entity data. The method can fully mine the hidden information of the knowledge graph data and improve the robustness by using the integrated model to combine the structure view and the feature view to classify the knowledge graph entity.

Description

TECHNICAL FIELD

[0001] The application provides a trusted knowledge graph network entity classification method based on a multi-view integrated model, mainly aims at the anti-adversarial defense of a knowledge graph entity classification model, improves the robustness of the knowledge graph entity classification model under adversarial attacks, and realizes accurate classification of knowledge graph entities. BACKGROUND

[0002] By observing data resources generated from complex systems in various industries, it can be found that there are often associated relationships between data contents or data. Mining and utilizing these associated relationships can understand the behavior patterns of groups and enhance intelligent service capabilities. A large number of research practices show that a graph (or a network) is a powerful tool for modeling data of different sources and different properties. It can very naturally abstractly represent various data resources through entity sets and associated relationships between the entity sets, thereby forming a large amount of graph data resources.

[0003] Among various graph data, a knowledge graph has become a hot spot of researchers. In the field of knowledge graphs, there are steps such as named entity recognition, relationship extraction, and triple construction. As a key pre-step in many downstream tasks, named entity recognition has been widely applied to scenarios such as machine translation, question answering, and knowledge graph construction. However, like other neural network models, entity classification models can also be attacked. Due to the discrete nature and connectivity of graphs, more challenges are brought to the data security of graphs.

[0004] In an open and adversarial application environment, an attacker can maliciously attack a model at data input, preprocessing, model learning, and data output of a knowledge graph intelligent system, thereby stealing training data and model parameters, reducing system performance, or even causing incorrect results. Especially in the related entity classification model in the knowledge graph, there are a large amount of sensitive data in these classification models, which are closely related to the privacy of users and involve the privacy security of users. However, the current graph neural network knowledge graph entity classification model algorithm in an open environment generally faces security risks. For example, in a knowledge graph model, an attacker can add entities or links maliciously to cause the knowledge graph model to recognize incorrectly. In addition, in a knowledge question and answer system, an attacker can infer the identity of a user according to a model output. Therefore, in a knowledge graph, how to defend against adversarial attacks on graph neural network models, improve the robustness of graph neural network models, and realize correct classification of entities in the knowledge graph is crucial. SUMMARY

[0005] Based on the idea and problems of the above method, combining the advantages and disadvantages of the existing method, the application provides a trusted knowledge graph entity classification method based on a multi-view integrated model, which can fully mine entity information in the knowledge graph entity network, improve the robustness of the graph neural knowledge graph entity recognition model under the attack, and improve the classification accuracy of the knowledge graph entity data.

[0006] The trusted knowledge graph entity classification method based on the multi-view integrated model comprises the following steps:

[0007] Obtain knowledge graph entity data in a knowledge graph entity network, and generate an original view of the knowledge graph entity data; the knowledge graph entity data comprises knowledge graph entity attribute information and knowledge graph entity association information;

[0008] Calculate a knowledge graph entity adjacency matrix according to the knowledge graph entity attribute information, and calculate a knowledge graph entity feature matrix according to the knowledge graph entity association relationship;

[0009] Construct a structure view for the knowledge graph entity adjacency matrix, and construct a feature view for the knowledge graph entity feature matrix;

[0010] Input the structure view and the original view into a structure-based model to output a first feature; input the feature view and the original view into a feature-based model to output a second feature;

[0011] Input the first feature and the second feature into a linear layer to integrate and output the category of the knowledge graph entity data.

[0012] The application has the following advantages:

[0013] (1) The application combines the models of multiple methods to analyze the knowledge graph entity citation data, and the classification recognition rate of the knowledge graph entity data is higher under the attack, because the method uses an integrated model to combine the structure view of the knowledge graph entity and the feature view of the knowledge graph entity to classify the knowledge graph citation entity, which can more fully mine the hidden information of the knowledge graph entity citation data and improve the robustness.

[0014] (2) The application combines the models of two methods to analyze the knowledge graph entity citation data, which can be applied to knowledge graph entity citation data without a feature matrix, because the model uses a structure similarity score to extract structure information, and the structure similarity score does not depend on the knowledge graph entity feature matrix, so that the high-order field information of the knowledge graph entity data can be extracted even if only the knowledge graph entity adjacency matrix is available, and the problem of being unable to use under the absence of the knowledge graph entity feature matrix can be solved. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 is a trusted knowledge graph entity classification method based on a multi-view integrated model of the present application;

[0016] Figure 2 is a trusted knowledge graph entity classification architecture based on a multi-view integrated model of the present application;

[0017] Figure 3 is a structural base model architecture used in the present application;

[0018] Figure 4 is a feature-based model architecture used in the present application; DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0020] It should be noted that the terms "include" and "have" and any variations thereof in the specification and claims of the present application and the above-described drawings are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units that are not clearly listed or inherent to the process, method, product or device.

[0021] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the drawings and in combination with the embodiments.

[0022] A knowledge graph is essentially a semantic network knowledge base. As one of the branches of artificial intelligence, a knowledge graph is a technology for complex knowledge modeling and potential relationship reasoning using a graph structure, which describes types, entities and their relationships in the objective world in a structured form. An entity is a thing in the objective world, and a type is an abstraction and generalization of things with the same attributes. A knowledge graph is an excellent way to associate citation data entities and infer implicit relationships.

[0023] Figure 1 is a trusted knowledge graph entity classification method based on a multi-view integrated model of the present application, as shown in Figure 1 the method comprises the following steps:

[0024] 101、Obtaining knowledge graph entity data in a knowledge graph entity network, and generating an original view of the knowledge graph entity data; the knowledge graph entity data includes knowledge graph entity attribute information and knowledge graph entity association information;

[0025] In the embodiment of the application, the knowledge graph entity attribute information includes knowledge graph entity identity information (entity ID), entity label (entity tag), and a set of word bags contained in the entity; and the entity association information includes entity association relationship.

[0026] In the embodiment of the application, the original view of the knowledge graph entity data can be represented as G=(A,X), wherein A=(V,E) is an adjacency matrix of a knowledge graph entity graph, V is a knowledge graph entity node (including identity information), E is an edge in a knowledge graph entity association graph, representing an association relationship of the knowledge graph entity, and X is feature information of the knowledge graph entity, such as a set of word bags contained in the knowledge graph entity; wherein the knowledge graph entity node can be represented by entity identity information (entity ID), the entity association relationship is a connection in the original view, the other end of the connection is another entity node, and the set of word bags contained in the entity is the feature information.

[0027] In the embodiment of the application, the knowledge graph entity network can be an existing graph citation network dataset, such as the Cora, Pubmed, and Citeseer datasets. For example, the Cora dataset includes 2708 graph nodes, 5429 edges, 1433 features, and 7 labels. The Citeseer dataset includes 3327 graph nodes, 4732 edges, 3703 features, and 6 labels. The Pubmed dataset includes 19717 graph nodes, 44338 edges, 500 features, and 3 labels. In the graph citation network, each node represents a paper, and the edges between the nodes represent the citation relationship between the papers. The features of the nodes are bag-of-words descriptions about the papers, and the categories of the nodes represent different types of the papers; and the knowledge graph entity citation data represents the citation relationship between the entity papers through a knowledge graph. In the knowledge graph entity data, the citation entity node-citation relationship-citation entity node can be regarded as a record, the first citation entity node is regarded as the subject, the citation relationship is regarded as the predicate, and the second citation entity node is regarded as the object, and the subject-predicate-object constitutes a record, and a plurality of records constitute the knowledge graph data in the embodiment of the application.

[0028] In some other embodiments of the present application, the knowledge graph entity citation data can be obtained from the citeulike-a dataset, i.e., the knowledge graph entity citation attribute information and the knowledge graph entity citation association information are obtained from the citeulike-a dataset. The citeulike-a dataset comes from the citeulike website, which is a website providing online literature management services and can help users manage literature of interest. The citeulike-a dataset contains 5509 users, 13519 articles, 5138 tags and 164863 user-article interaction records, in addition to which the dataset also contains paper citation data and paper keyword data; the paper citation data and the paper keyword data are contained in the knowledge graph entity attribute information and the knowledge graph association information.

[0029] Figure 2 is a schematic diagram of a trusted knowledge graph entity classification method based on a multi-view integrated model according to the present application, as shown in Figure 2 According to the present application, the knowledge graph entity adjacency matrix and the knowledge graph entity feature matrix are processed respectively to form the corresponding structure view and the feature view, the structure view and the feature view are processed respectively by using the corresponding structure base model and the feature base model, and finally the corresponding features are output by using the integrated model, so that the knowledge graph entity nodes can be classified in combination with the information of the structure view of the knowledge graph entity nodes and the feature view of the knowledge graph entity, which can more fully mine the hidden information of the knowledge graph entity data and improve the robustness.

[0030] 102. Calculate the knowledge graph entity feature matrix according to the knowledge graph entity attribute information, and calculate the knowledge graph entity adjacency matrix according to the knowledge graph entity association relationship;

[0031] In the embodiments of the present application, the knowledge graph entity adjacency matrix A is determined by the entity association relationship, and the knowledge graph entity feature matrix X is determined by the entity feature information, and the specific determination manner can refer to the embodiments corresponding to step 101.

[0032] 103. Construct a structure view for the knowledge graph entity adjacency matrix, and construct a feature view for the knowledge graph entity feature matrix;

[0033] In the embodiments of the present application, constructing the structure view for the knowledge graph entity adjacency matrix includes calculating the similarity scores between the knowledge graph entity nodes; determining the K nearest neighbor knowledge graph entity nodes of the current knowledge graph entity node according to the similarity scores between the knowledge graph entity nodes, and constructing a K nearest neighbor structure graph, wherein the calculation formula of the similarity scores between the knowledge graph entity nodes is:

[0034]

[0035] wherein, simrank(a, b) represents the similarity score between the knowledge graph entity node a and the knowledge graph entity node b, I(a) and I(b) represent the knowledge graph entity node set pointing to the knowledge graph entity node a and the knowledge graph entity node b respectively, and the damping coefficient C is a constant between 0 and 1; I i (a) represents the i-th entity node in the neighbor set pointing to the a knowledge graph entity node, and I(b) represents the j-th entity node in the neighbor set pointing to the b knowledge graph entity node.

[0036] In the embodiment of the application, the feature view is constructed by calculating the cosine similarity between each knowledge graph entity feature vector, determining the K nearest neighbor knowledge graph entity feature vectors of the current knowledge graph entity feature vector according to the cosine similarity between each knowledge graph entity feature vector, and constructing a K nearest neighbor structure graph, wherein the calculation formula of the cosine similarity between the knowledge graph entity feature vectors is as follows:

[0037]

[0038] wherein, cosinesimilarity(A, B) represents the cosine similarity between the knowledge graph entity feature vector A and the knowledge graph entity feature vector B, A i and B i represent the i-th component of the knowledge graph entity feature vector A and the knowledge graph entity feature vector B respectively; and n represents the number of components of the knowledge graph entity feature vector.

[0039] In the embodiment, two different ways are adopted when constructing the view, which can be applied to the knowledge graph entity citation data without a feature matrix. Since the model uses the structural similarity score to extract the structural information, the structural similarity score does not depend on the knowledge graph entity feature matrix, and the high-order field information of the knowledge graph entity citation data can still be extracted when only the knowledge graph entity adjacency matrix is available, thereby solving the problem that the model cannot be used when the knowledge graph entity feature matrix is missing.

[0040] 104, input the structure view and the original view into the structure-based model to output a first feature, and input the feature view and the original view into the feature-based model to output a second feature;

[0041] In the embodiment of the application, Figure 3 is the structure-based model architecture diagram adopted by the application; as Figure 3As shown, the structure-based model uses the original view and the structure view as input, uses MLP to adaptively combine the information of the original view and the structure view, and generates an output first feature through two layers of graph convolutional neural network; the feature-based model uses the original view and the feature view as input, uses MLP to adaptively combine the information of the original view and the feature view, and generates an output second feature through two layers of graph convolutional neural network.

[0042] In some embodiments, the structure-based model includes using two layers of MLP layers to learn the scoring coefficients between the original view and the structure view; combining the knowledge graph entity adjacency matrix to calculate the combination coefficients between the original view and the structure view; using a linear layer to learn the proportion of each knowledge graph entity node adding a self-loop; using two layers of GCN network to encode and embed the knowledge graph entity node information, processing the combination coefficients and the proportion of each knowledge graph entity node adding a self-loop, and calculating a first feature. The specific structure of the structure-based model is as follows:

[0043] MLP is used to learn the combination coefficients of the original view and the structure feature view; in the structure-based model, the aggregation process can be described as:

[0044]

[0045] wherein, represents the combination graph of the lth layer of the knowledge graph entity recognition system structure-based model, represents the degree matrix of the knowledge graph original view, represents the adjacency matrix of the knowledge graph original view, D stru represents the degree matrix of the knowledge graph structure KNN view, A stru represents the adjacency matrix of the knowledge graph structure KNN view, which is calculated by similarity score; represents the scoring coefficients of the original view and the structure view learned by the lth MLP layer. We use MLP to learn the combination coefficients

[0046]

[0047] wherein, H (l-1) is the hidden representation of the last layer of the knowledge graph entity recognition system structure-based model, H (0) = X, and represent the weight coefficient and the bias of the MLP lth layer, respectively, and σ is an activation function.

[0048] The linear layer is used to learn the proportion of each knowledge graph entity node feature adding a self-loop to the combination graph of the structure-based model .

[0049] We model the proportion of each self-loop with a learnable orthogonal matrix

[0050] We denote the proportion of each knowledge graph entity node i in the l-th linear layer as is denoted as:

[0051]

[0052] where H (l-1) is the hidden representation of the previous layer, H (0) = X, and represent the weight coefficient and bias of the l-th layer, respectively.

[0053] Two-layer GCN is used to encode and embed the information of knowledge graph entity nodes. In the above steps, we can model the combination view as:

[0054]

[0055] Therefore, the hidden representation of each layer in the GCN can be represented as:

[0056]

[0057] where H is the hidden representation of the previous layer, is the learnable parameter of the l-th layer GCN in the structure-based model, and σ is the activation function.

[0058] In some embodiments, Figure 4 is the feature-based model architecture used in the present application, as shown in Figure 4 The feature-based model includes using two layers of MLP to learn the score coefficient between the original view and the feature view; the combination coefficient between the original view and the feature view is calculated by combining the knowledge graph entity feature matrix; the proportion of each knowledge graph entity node adding self-loop is learned by using linear layer; the information of knowledge graph entity nodes is encoded and embedded by using two-layer GCN network, and the combination coefficient and the proportion of each knowledge graph entity node adding self-loop are processed to calculate the second feature. The specific structure of the feature-based model is as follows:

[0059] MLP is used to learn the combination coefficient of the original view and the feature view; in the feature-based model, the aggregation process can be described as:

[0060]

[0061] where, represents the combination graph of the l-th layer knowledge graph entity recognition system feature-based model, The degree matrix represents the original view of the knowledge graph. D represents the adjacency matrix of the original view of the knowledge graph. f A represents the degree matrix of the feature KNN view of the knowledge graph. f The adjacency matrix of the feature KNN view of the knowledge graph is calculated from the cosine similarity. This represents the score coefficient learned by the l-th MLP layer, which combines the original view and the feature view. To combine the scoring coefficients of the original view and the feature view, we use a two-layer MLP to learn...

[0062]

[0063] Among them, H (l-1) H is the hidden representation of the feature base model of the upper-layer knowledge graph entity recognition system. (0) =X, and σ represents the weight coefficients and bias of the l-th layer of the MLP, respectively, and σ is the activation function.

[0064] Linear layers are used to learn the proportion of self-loops added to each node;

[0065] We use a learnable orthogonal matrix to model the weight of each self-loop.

[0066] We will determine the weight of each knowledge graph entity node i in the l-th linear layer. Represented as:

[0067]

[0068] Among them, H (l-1) H is the hidden representation of the previous layer. (0) =X, and These represent the weight coefficient and bias of the l-th layer, respectively.

[0069] Two layers of GCN are used to encode and embed node information. In the above steps, we can model the combined graph as follows:

[0070]

[0071] Therefore, the hidden representation of each layer in GCN can be represented as:

[0072]

[0073] in This is a hidden representation of the layer above. Let σ be the learnable parameters of the l-th layer GCN in the feature base model of the knowledge graph entity recognition system, and σ be the activation function of the GCN. View the knowledge graph entity network structure A. stru The adjacency matrix A of the original view of the knowledge graph entity network is input into the structural base model to produce the output. f The adjacency matrix A of the original view in the knowledge graph entity network is input into the feature base model to produce the output. The structural base model has the same architecture as the feature base model. Both base models consist of a graph convolutional neural network trained to generate vector-like input and output copies of data attributes. To calculate the weights of each node combining the original graph and the view, each base model also includes an MLP to calculate the weights of combining the original view with the corresponding view. In each base model, the input data is processed by the MLP to calculate the combination rate, and then passed through two layers of graph convolutional neural networks to produce the output. The structural feature embedding information generated by the structural base model, i.e., the first feature, corresponds to Z. stru The node feature embedding information generated by the feature basis model, i.e., the second feature, corresponds to Z. f .

[0074] 105. Input the first feature and the second feature into the linear layer and integrate the output categories of the knowledge graph entity citation data.

[0075] In this embodiment of the invention, the output Z generated by the structural basis model is... stru The output Z generated by the feature basis model f As input, it is fed into a linear layer, which adaptively produces the final output, which is then used to predict the category of the graph nodes.

[0076] Specifically, an attention coefficient is defined for each knowledge graph entity node from the adaptive ensemble, where the attention coefficient of knowledge graph entity node i in model s is... The model is as follows:

[0077]

[0078] Where σ is the activation function, W i and b i These are the parameters and biases of the linear layer, respectively.

[0079] The final output feature Z of entity node i in the knowledge graph i Represented as:

[0080]

[0081] In the above process, the corresponding Loss function can be defined as:

[0082]

[0083] wherein, D L is a labeled knowledge graph entity node, y i is a classification label of the knowledge graph entity node v i , for example, the classification result is one of the following six categories: Agents (agent), AI (artificial intelligence), DB (database), IR (information retrieval), ML (machine learning) and HCI (human-computer interaction), and l(·,·) is a loss function for measuring the degree of difference between the predicted value and the true label. By using the method proposed in the application, the existing graph citation network dataset is input into the model proposed in the application. The experimental results are shown in Table 1.

[0084] Table 1 Entity classification recognition accuracy under Mettack attack

[0085]

[0086] The trusted knowledge graph entity classification method based on the multi-view integrated model proposed in the application not only overcomes the problem that the existing knowledge graph entity citation classification model can only be applied to knowledge graph entity citation data with a feature matrix, but also has a higher recognition accuracy than other graph neural network entity recognition models.

[0087] A person of ordinary skill in the art can understand that all or part of the steps in the above-mentioned embodiments can be completed by a program instructing related hardware, and the program can be stored in a computer readable storage medium, which can include ROM, RAM, magnetic disk or optical disk, etc.

[0088] Although the embodiments of the application have been shown and described, it can be understood by those of ordinary skill in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the application, and the scope of the application is defined by the appended claims and their equivalents.

Claims

1. A method for trusted knowledge graph entity classification based on a multi-view integrated model, characterized in that, The method comprises the following steps: Obtain knowledge graph data in a knowledge graph network, and generate an original view of the knowledge graph data; the knowledge graph data comprises knowledge graph entity attribute information and knowledge graph entity association information; Calculate a knowledge graph entity feature matrix according to the knowledge graph entity attribute information, and calculate a knowledge graph entity adjacency matrix according to the knowledge graph entity association relationship; Construct a structure view from the knowledge graph entity adjacency matrix, and construct a feature view from the knowledge graph entity feature matrix; Input the structure view and the original view into a structure-based model to output a first feature, and input the feature view and the original view into a feature-based model to output a second feature; the structure-based model uses the original view and the structure view as input, uses MLP to adaptively combine the information of the original view and the structure view, and generates the first feature through two layers of graph convolutional neural networks; the feature-based model uses the original view and the feature view as input, uses MLP to adaptively combine the information of the original view and the feature view, and generates the second feature through two layers of graph convolutional neural networks; Input the first feature and the second feature into a linear layer to integrate and output the category of the knowledge graph entity. 2.The method of claim 1, wherein, The knowledge graph entity attribute information comprises knowledge graph entity identity information, labels, and a set of word packages contained in the entity; the knowledge graph entity association information comprises entity association relationships. 3.The method of claim 1, wherein, Constructing the structure view from the knowledge graph entity adjacency matrix comprises calculating the similarity scores between each knowledge graph entity; according to the similarity scores between each knowledge graph entity, the K nearest neighbor knowledge graph entities of the current knowledge graph entity are determined, and a K nearest neighbor structure graph is constructed, wherein the calculation formula of the similarity scores between knowledge graph entity nodes is: ； wherein, represents a knowledge graph entity node represents a similarity score between a knowledge graph entity node is a constant value between 0 and 1; represents the i-th entity node in the neighbor set pointing to a knowledge graph entity node represents the j-th entity node in the neighbor set pointing to a knowledge graph entity node​​​​​ 4.The method of claim 1, wherein, Constructing the feature view from the knowledge graph entity feature matrix comprises calculating the cosine similarity between each knowledge graph entity feature vector; according to the cosine similarity between each knowledge graph entity feature vector, the K nearest neighbor knowledge graph entity feature vectors of the current knowledge graph entity feature vector are determined, and a K nearest neighbor feature graph is constructed, wherein the calculation formula of the cosine similarity between knowledge graph entity feature vectors is: ； in, Represents the entity feature vector of a knowledge graph With knowledge graph entity feature vectors Cosine similarity between them and These represent the entity feature vectors of the knowledge graph. and knowledge graph entity feature vectors The One component; This represents the number of components in the feature vector of an entity in a knowledge graph.

5. The method of claim 1, wherein, The structure-based model comprises using two layers of MLP layers to learn the scoring coefficients between the original view and the structure view; combining the knowledge graph entity adjacency matrix to calculate the combination coefficients between the original view and the structure view; Using a linear layer to learn the proportion of each knowledge graph entity node adding a self-loop; using two layers of GCN networks to encode and embed the knowledge graph entity information, processing the combination coefficients and the proportion of each knowledge graph entity node adding a self-loop, and calculating the first feature. 6.The method of claim 1, wherein, The feature-based model comprises using two layers of MLP layers to learn the scoring coefficients between the original view and the feature view; combining the knowledge graph entity feature matrix to calculate the combination coefficients between the original view and the feature view; The proportion of each knowledge graph entity node adding a self-loop is learned by using a linear layer; information of the knowledge graph entity node is encoded and embedded by using a two-layer GCN network, and the combination coefficient and the proportion of each knowledge graph entity node adding a self-loop are processed to calculate a second feature. 7.The method of claim 1, wherein, The category of the integrated output knowledge graph entity includes a first feature output by a structure-based model and a second feature output by a feature-based model which are adaptively combined by using a linear layer; specifically, an attention coefficient is defined for each knowledge graph entity node to adaptively integrate, the combination coefficient of the knowledge graph entity node is adaptively weighted with the corresponding feature to calculate a final output feature, and the category of the knowledge graph entity data is determined according to the final output feature.

Images