A method for embedding inductive reasoning into knowledge graph of market dishonesty information

The super-relational triple features are extracted through BERT and convolutional neural networks, combined with meta-learning technology, and the problem that traditional knowledge graph embedding methods cannot handle invisible entities and relationships is solved, and the accurate embedding and generalization of knowledge graphs for market breach of trust information is achieved. The ability to improve the performance of link prediction and knowledge graph fusion.

CN118886498BActive Publication Date: 2024-11-29湖南工商大学
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411394827.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-10-08
Publication Date
2024-11-29
Estimated Expiration
2044-10-08

AI Technical Summary

Technical Problem

Traditional knowledge graph embedding methods cannot effectively embed entities and relationships that are not visible during training, cannot accurately model hyperrelational triples in market breach of trust information, and cannot link predictions for entities and relationships that are not visible in the knowledge graph.

Method used

The BERT pre-trained model is used to extract key-value pair features of hyperrelational triplets, combine convolutional neural networks to fusion auxiliary information word vectors and relationship vectors, build comprehensive relationship embeddings, and divide the knowledge graph into sub-graph tasks through meta-learning for training , generate entity and relation embeddings.

Benefits of technology

The accurate embedding of invisible entities and relationships in the market's untrustworthy information knowledge graph is achieved, the ability to represent relationship embedding is enhanced, and the ability to generalize to other untrustworthy information knowledge graphs is reduced, and the task time cost is reduced.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118886498B_ABST
    Figure CN118886498B_ABST
Patent Text Reader

Abstract

The present application relates to a method for inductive reasoning embedding of a market dishonesty information knowledge graph, the method comprising: obtaining a market dishonesty information knowledge graph as a source knowledge graph, and using a BERT pre-trained model to perform feature extraction on key-value pairs of super-relation triples in the source knowledge graph to obtain auxiliary information word vectors; using a convolutional neural network to fuse the auxiliary information word vectors with the constructed relationship vectors, and performing correction constraints on the fusion results to obtain comprehensive relationship embeddings; constructing type-level information embeddings of entities according to the relationship connections of entities in the source knowledge graph; constructing instance-level information embeddings of entities according to the type-level information embeddings of entities and the neighborhood information of the corresponding entities; performing meta-learning based on the source knowledge graph, the comprehensive relationship embeddings, and the instance-level information embeddings of entities to obtain a knowledge graph embedding model; inputting the target knowledge graph to be processed into the knowledge graph embedding model to obtain entity embeddings and relationship embeddings.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of knowledge graph inductive reasoning embedding, and in particular to a method for embedding market dishonesty information knowledge graph inductive reasoning embedding. Background Art

[0002] The knowledge graph of market dishonest information is an intelligent technology that uses the knowledge structure of the knowledge graph to manage market dishonest information. It studies the internal relationships of dishonest entities such as enterprises and departments through information fusion embedding. A large number of triples carry additional information and belong to the category of super-relational triples. By introducing the knowledge graph embedding model of inductive reasoning, it can learn relational semantics independent of entities and has strong generalization ability.

[0003] However, traditional knowledge graph embedding methods can only generate embeddings for visible entities and relations during training that can be used for downstream tasks, and cannot generate effective embeddings for invisible entities during training. Existing inductive reasoning embedding methods can train a knowledge graph embedding model on a set of entities, so that it can be generalized to the embedding task of invisible entities, but cannot represent any information other than entities and relations, that is, it cannot accurately model hyper-relational triples with a large amount of auxiliary information such as market default information. Existing hyper-relational knowledge graph embedding methods can solve the embedding problem of hyper-relational triples and perform downstream tasks such as link prediction while retaining the original structural information of the knowledge graph based on triples, but they can only model known entities and relations in the knowledge graph, and cannot perform downstream tasks such as link prediction for invisible entities and relations in the knowledge graph. Summary of the invention

[0004] Based on this, it is necessary to provide a method for embedding market dishonesty information knowledge graph inductive reasoning based on hyper-relation triples, which includes:

[0005] S1: Obtain a market dishonesty information knowledge graph as a source knowledge graph, and use a BERT pre-trained model to extract features from key-value pairs of super-relation triples in the source knowledge graph to obtain auxiliary information word vectors;

[0006] S2: using a convolutional neural network to fuse the auxiliary information word vector with the constructed relationship vector, and performing correction constraints on the fusion result to obtain a comprehensive relationship embedding;

[0007] S3: constructing entity type-level information embedding according to the relational connections of entities in the source knowledge graph;

[0008] S4: constructing instance-level information embedding of the entity according to the type-level information embedding of the entity and the neighborhood information of the corresponding entity;

[0009] S5: Perform meta-learning based on the source knowledge graph, the comprehensive relationship embedding and the instance-level information embedding of the entity to obtain a knowledge graph embedding model; input the target knowledge graph to be processed into the knowledge graph embedding model to obtain entity embedding and relationship embedding.

[0010] Beneficial effects: This method uses the key-value pairs in the hyper-relation triples, converts them into auxiliary information and merges them with the basic relation embedding to enhance the representation ability of the relation embedding; uses the relational connection of entities to initialize entities, and further embeds instance-level information of entities to form an entity embedding learning method for inductive reasoning; uses the idea of ​​meta-learning to divide the knowledge graph into multiple sub-graph tasks for meta-task training, so that the knowledge graph embedding model has the ability to generalize to invisible entities. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0012] Figure 1 This is a flowchart of the market dishonesty information knowledge graph inductive reasoning embedding method of the embodiment of the present application. DETAILED DESCRIPTION

[0013] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are described in detail below in conjunction with the accompanying drawings. In the following description, many specific details are set forth to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar improvements without violating the connotation of the present application, so the present application is not limited by the specific embodiments disclosed below.

[0014] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of this application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0015] The purpose of this embodiment is to solve the embedding accuracy problem of the knowledge graph of dishonest information, optimize the three defects of the above-mentioned background technology, and be able to learn the representation of entities, relations and auxiliary information (k, v) in the hyper-relation triple of dishonest information, while retaining the original structure of the knowledge graph formed by (h, r, t), and be able to generate accurate embeddings for new entities that are not visible during training. After embedding a knowledge graph of dishonest information, this method can generate embeddings for newly added entities in the knowledge graph without retraining, and even has the ability to generalize to other knowledge graphs of dishonest information, greatly reducing the time cost of related tasks.

[0016] like Figure 1 As shown, this embodiment provides a method for embedding market dishonesty information knowledge graph inductive reasoning, the method comprising:

[0017] S1: Obtain a market dishonesty information knowledge graph as a source knowledge graph, and use a BERT pre-trained model to extract features from key-value pairs of hyper-relation triples in the source knowledge graph to obtain auxiliary information word vectors.

[0018] The basic data unit in the knowledge graph (KG) is a triple, each triple is represented by a head entity, a relation, and a tail entity, and is generally represented as (h, r, t). However, this representation simplifies the complexity of the data stored in the knowledge graph, especially for super-relational facts. In addition to the basic triple (h, r, t), each super-relational triple also contains multiple related key-value pairs (k, v), where k represents the super-relational name and v represents the super-relational content. Conventional embedding models will ignore this additional information, resulting in missing knowledge graph information, which in turn affects the embedding accuracy.

[0019] The basic data units in the knowledge graph of dishonest information store various credit record information between enterprises and departments. It is obviously impossible to fully explain this type of information with simple relationships such as "default" and "warning", so the triple of dishonest information carries a lot of additional information, including specific content such as the amount of default and the warning level. Therefore, the basic data units in the knowledge graph of dishonest information are mostly hyper-relation triples, and these additional information need to be modeled and represented.

[0020] This embodiment regards the additional information key-value pairs (k, v) in the super-relation triples of the default information as auxiliary information to enhance the representation learning performance of the knowledge graph, integrates them into the relation embedding vector, and uses an inductive reasoning embedding method based on meta-learning to enable the knowledge graph embedding model to accurately generate entity embeddings when facing new entities.

[0021] Specifically, the key-value pairs of the super relation triples are input into the BERT pre-training model for feature extraction, and the first auxiliary information word vector K and the second auxiliary information word vector V are output.

[0022] S2: A convolutional neural network is used to fuse the auxiliary information word vector with the constructed relationship vector, and correction constraints are performed on the fusion result to obtain a comprehensive relationship embedding.

[0023] Specifically, a randomly initialized basic relationship matrix is ​​constructed, wherein the basic relationship matrix includes relationship vectors of multiple basic relationships;

[0024] In each super-relation triple, a convolutional neural network is used to fuse all the first auxiliary information word vectors K and the second auxiliary information word vectors V with each of the relationship vectors to obtain a relationship embedding of the fused auxiliary information;

[0025] Furthermore, the process of obtaining the relation embedding integrating the auxiliary information includes:

[0026] Step 1: Use zero padding to unify the dimensions of the relationship vector, the first auxiliary information word vector K, and the second auxiliary information word vector V into the dimension of the relationship vector;

[0027] Step 2: Perform 2D plane convolution on the relation vector and the first auxiliary information word vector K and the second auxiliary information word vector V of each key-value pair to obtain a new relation embedding;

[0028] In this embodiment, the convolution kernel size of the convolutional neural network is set to 3*3, and the convolution kernel step size is set to 1.

[0029] Step 3: Repeat step 2 until all the relationship vectors are traversed to obtain multiple new relationship embeddings;

[0030] Step 4: Perform weighted averaging on all the new relation embeddings to obtain the relation embedding of the fused auxiliary information. The calculation formula is as follows:

[0031] ;

[0032] in, Indicates w The super relation triples are composed of r The relation embedding of fused auxiliary information obtained by the relation vector; n Indicates w The number of key-value pairs in a superrelation triple; Indicates w The first i The weight value of the group key-value pair; Indicates r The relationship vector isi New relational embeddings for group key-value pairs.

[0033] Obtaining relational embedding vectors fused with auxiliary information by convolutional neural networks , the information density of relation embedding is improved by introducing auxiliary information, but it also weakens the expression ability of relation embedding for basic relations r and triples (h, r, t). Therefore, certain correction methods are needed to constrain the way of fusing auxiliary information.

[0034] Embed the relationship obtained after convolution into a vector The idea of ​​contrastive learning is used for correction, so that the vectors constructed by the same relationship are closer in the vector space, further improving the performance of knowledge graph representation learning. Contrastive learning brings the relationship embedding constructed by the same relationship as the basic relationship closer in the vector space, and makes the relationship embedding of different basic relationships more evenly distributed. Specifically, the relationship embedding of the fused auxiliary information is corrected and constrained to obtain a comprehensive relationship embedding, including:

[0035] Step 1: Construct positive and negative samples, and transform the relationship vectors in the same training batch The obtained relation embedding fused with auxiliary information As a positive sample, the relationship vector The obtained relation embedding fused with auxiliary information As negative samples; relation vector For basic relationship The vector representation of the relationship vector For basic relationship The vector representation of In addition to the basic relationship Any other basic relationship;

[0036] Step 2: Input the positive and negative samples into a multi-layer perceptron for contrastive learning, and output the comprehensive relationship embedding.

[0037] The expression of contrastive learning loss is:

[0038] ;

[0039] ;

[0040] in, represents contrastive learning loss; n express The number of positive samples; Represents any relation embedding that incorporates auxiliary information; Indicates j positive samples; Indicates knegative samples; N represents the number of all super-relation triples in the batch; Indicates cosine similarity calculation; represents the temperature function.

[0041] The role of contrastive learning is to form a comprehensive relational embedding that can express auxiliary information while retaining basic relational information to the greatest extent. , this loss function adjusts the parameters of both MLP and convolutional neural network during training.

[0042] Completed construction of comprehensive relational embedding After that, the triple is composed of Updated to , the vector representation of the triple is given by ( ) is updated to ( ), which retains the original triplet information while incorporating auxiliary information.

[0043] S3: Constructing type-level information embedding of entities according to the relational connections of entities in the source knowledge graph.

[0044] The idea of ​​inductive learning needs to simulate the human cognitive process of new entities. The relationship around the entity is the most natural information, that is, the type of the entity is preliminarily judged by the relationship and relationship direction connected to the entity. Specifically, this step includes:

[0045] Step 1: Design a relation domain embedding and a relation range embedding for each basic relation; the relation domain embedding is used to transmit the hidden information of the basic relation when it is output by the head entity; the relation range embedding is used to transmit the hidden information of the basic relation when it is input by the tail entity;

[0046] Step 2: For any entity, the relationship domain embedding and the relationship range embedding that reflect the input-output relationship are weighted averaged to obtain the type-level information embedding of the entity; the calculation formula is:

[0047] ;

[0048] in, Indicates e Type-level information embedding of entities; express For e The weight of an entity; express For e The weight of an entity; Indicates r The relational domain embedding of basic relations; Indicates r The relational scope of each basic relation is embedded; Indicates e The total number of basic relations output by each entity; Indicates e The total number of basic relations entered for each entity.

[0049] S4: Constructing instance-level information embedding of the entity according to the type-level information embedding of the entity and the neighborhood information of the corresponding entity.

[0050] After completing the initialization of the entity, that is, after the type-level information of the entity is embedded, if there are two entities with the same relationship connection, we can only judge which entity they belong to through their type information, but cannot determine whether they are specific individuals. Therefore, it is also necessary to represent the instance-level information of the entity based on the neighborhood structure.

[0051] Specifically, the step includes:

[0052] Step 1: The neighborhood information of the entity includes the structural information of the multi-hop neighbors of the entity. The structural information and the type-level information of the entity are embedded and input into the graph neural network to obtain the hidden entity representation of the entity. l The update formula of the layer is:

[0053] ;

[0054] in, Indicates e The entity in l Hidden entity representation of the layer; ReLU activation function. Indicates all e The number of neighbor triplets that an entity is directly connected to; Indicates l Layer r The transformation matrix of the basic relations; Indicates l The self-circular transformation matrix of entities in the layer; Indicates e The entity in l- The hidden entity representation of the first layer; the type-level information of the entity is embedded as e The hidden entity representation of entities at layer 0; h Representation and Entity e Directly connected entities , Representing Entities h In the l- Hidden entity representation at layer 1.

[0055] Step 2: Use the skip knowledge matrix to concatenate with the hidden entity representation to obtain the instance-level information embedding of the entity; the skip knowledge matrix is ​​used to convert the concatenated hidden entity representation into an entity embedding; the calculation formula is:

[0056] ;

[0057] in, Indicates e Instance-level information embedding of entities; represents the jump knowledge matrix; Represents vector concatenation.

[0058] S5: Perform meta-learning based on the source knowledge graph, the comprehensive relationship embedding and the instance-level information embedding of the entity to obtain a knowledge graph embedding model; input the target knowledge graph to be processed into the knowledge graph embedding model to obtain entity embedding and relationship embedding.

[0059] In this example, the goal of using inductive reasoning is to make the above instance-level information embedding generation process And the generation process of comprehensive relational embedding It can output reasonable relation embedding and entity embedding for unseen super-relation triples, that is, learn to predict new entity embedding and relation embedding. Inspired by the concept of "learning to learn" in meta-learning, the embedding model is trained on a series of meta-learning tasks using meta-learning theory.

[0060] Specifically, the step includes:

[0061] Step 1: Use a random walk method to sample a group of sub-knowledge graphs from the market dishonesty information knowledge graph, and re-label the entities in each sub-knowledge graph as invisible entities;

[0062] Step 2: Split the hyperrelation triples in any processed sub-knowledge graph into query triples and support triples to construct a meta-learning task; the expression of the meta-learning task is:

[0063] ;

[0064] ;

[0065] in, Indicates i a meta-learning task; Represents the set of entities in the market dishonesty information knowledge graph; Represents the set of relations in the knowledge graph of market dishonesty information; Indicates i Support triplets for each meta-learning task; Indicates i query triples for a meta-learning task; s Represents the source knowledge graph; Represents the sub-knowledge graph sampled by the meta-task from the source knowledge graph; Represents the entity set of the source knowledge graph; Represents the relation set of the source knowledge graph.

[0066] Step 3: In each meta-learning task, an entity embedding is obtained according to the support triples, and a loss function is calculated according to the query triples, the comprehensive relation embedding, and the instance-level information embedding of the entity; the expression of the loss function is:

[0067] ;

[0068] ;

[0069] ;

[0070] in, represents the loss function; represents a positive triple; h represents the head entity of the positive triple; Indicates the relationship of fusion auxiliary information; t represents the tail entity of the three positive tuples; in order to speed up the fitting speed of the model, the concept of negative samples is also introduced in the meta-knowledge training process, that is, in the positive triples Randomly replace the head entity h or the tail entity t; Indicates i negative triples; Indicates i The head entity of a negative triple; Indicates i The tail entity of a negative triple; Indicates i query triples for a meta-learning task; Represents the sigmoid function; represents the marginal bias; represents the positive triple scoring function; represents the negative triple scoring function; Represents the self-adversarial weight value of the negative triple; represents the temperature coefficient of negative triplet sampling; Represents the instance-level information embedding of the head entity; represents comprehensive relational embedding; Represents instance-level information embedding of the tail entity; k Represents the number of negative triplets.

[0071] Step 4: constructing a meta-training objective of the meta-learning task based on the entity embedding and the loss function;

[0072] The expression of the meta-training objective is:

[0073] ;

[0074] ;

[0075] in, represents the first learning parameter; represents the second learning parameter; represents the third learning parameter; represents the fourth learning parameter; represents the comprehensive relation embedding matrix in the current task; m represents the number of meta-learning tasks; Indicates i The training loss for each meta-learning task; Represents the generation process of embedding instance-level information in query triples And the generation process of comprehensive relational embedding Generation losses; Indicates i Contrastive learning loss in a meta-learning task; Indicates i Loss function in a meta-learning task.

[0076] Step 5: Train the meta-learning task based on the meta-training objective to obtain the knowledge graph embedding model;

[0077] Meta-knowledge learning is obtained by meta-training the source KG through meta-learning tasks. After meta-training, it can be used , The invisible relations and entities during training are embedded and can be generalized to the target KG to achieve inductive knowledge graph embedding. Note that the invisible relations here refer to the new relations generated by the super-relation triples under the condition of the existing basic relations.

[0078] First, the parameter freezing method is adopted, that is, freezing , Parameters of, and all relationship vectors R r and comprehensive relational embedding .

[0079] Subsequently, the above knowledge graph embedding model is regarded as an entity embedding and relationship embedding generator, and applied to the target KG for knowledge graph embedding and downstream tasks.

[0080] The market dishonesty information knowledge graph inductive reasoning embedding method based on hyper-relation triples provided in this embodiment utilizes the key-value pairs in the hyper-relation triples, converts them into auxiliary information and merges them with the basic relation embedding, thereby enhancing the representation capability of relation embedding; utilizes the relational connection of entities to initialize entities, further embeds instance-level information of entities, and forms an entity embedding learning method for inductive reasoning; utilizes the idea of ​​meta-learning to divide the knowledge graph into multiple sub-graph tasks for meta-task training, thereby enabling the knowledge graph embedding model to have the ability to generalize to invisible entities.

[0081] The market dishonesty information knowledge graph inductive reasoning embedding method based on hyper-relation triples provided in this embodiment is based on the dishonesty information knowledge graph inductive reasoning embedding model based on auxiliary information, which solves the problem that the traditional knowledge graph embedding model cannot generate accurate embedding for invisible entities during model training. At the same time, it is of great significance for improving the performance of a series of internal tasks and external tasks such as link prediction and knowledge graph fusion.

[0082] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0083] The above-described embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be construed as limiting the scope of the patent application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent application shall be subject to the attached claims.

Claims

1. A method for embedding market dishonesty information knowledge graph inductive reasoning, characterized in that: include: S1: Obtain a market dishonesty information knowledge graph as a source knowledge graph, and use a BERT pre-trained model to extract features from key-value pairs of super-relation triples in the source knowledge graph to obtain auxiliary information word vectors; Input the key-value pair of the super relation triple into the BERT pre-training model for feature extraction, and output the first auxiliary information word vector K and the second auxiliary information word vector V; S2: Using a convolutional neural network to fuse the auxiliary information word vector with the constructed relationship vector, correcting and constraining the fusion result to obtain a comprehensive relationship embedding, including: Constructing a randomly initialized basic relationship matrix, wherein the basic relationship matrix includes relationship vectors of multiple basic relationships; In each super-relation triple, a convolutional neural network is used to fuse all the first auxiliary information word vectors K and the second auxiliary information word vectors V with each of the relationship vectors to obtain a relationship embedding of the fused auxiliary information; The process of obtaining the relation embedding that incorporates the auxiliary information includes: Step 1: Use zero padding to unify the dimensions of the relationship vector, the first auxiliary information word vector K, and the second auxiliary information word vector V into the dimension of the relationship vector; Step 2: Perform 2D plane convolution on the relation vector and the first auxiliary information word vector K and the second auxiliary information word vector V of each key-value pair to obtain a new relation embedding; Step 3: Repeat step 2 until all the relationship vectors are traversed to obtain multiple new relationship embeddings; Step 4: Perform weighted averaging on all the new relation embeddings to obtain the relation embedding of the fused auxiliary information. The calculation formula is as follows: ; in, Indicates w The super relation triples are composed of r The relation embedding of fused auxiliary information obtained by the relation vector; n Indicates w The number of key-value pairs in a superrelation triple; Indicates w The first i The weight value of the group key-value pair; Indicates r The relationship vector is i New relational embeddings for group key-value pairs; Correcting constraints are applied to the relation embedding of the fused auxiliary information to obtain a comprehensive relation embedding, including: Step 1: Construct positive and negative samples, and transform the relationship vectors in the same training batch The obtained relation embedding fused with auxiliary information As a positive sample, the relationship vector The obtained relation embedding fused with auxiliary information As negative samples; relation vector For basic relationship The vector representation of the relationship vector For basic relationship The vector representation of In addition to the basic relationship Any other basic relationship; Step 2: Input the positive and negative samples together into a multi-layer perceptron for comparative learning, and output the comprehensive relationship embedding; S3: constructing entity type-level information embedding according to the relational connections of entities in the source knowledge graph; S4: constructing instance-level information embedding of the entity according to the type-level information embedding of the entity and the neighborhood information of the corresponding entity; S5: Perform meta-learning based on the source knowledge graph, the comprehensive relationship embedding and the instance-level information embedding of the entity to obtain a knowledge graph embedding model; input the target knowledge graph to be processed into the knowledge graph embedding model to obtain entity embedding and relationship embedding.

2. The method for embedding market dishonesty information knowledge graph inductive reasoning according to claim 1 is characterized in that: Constructing the type-level information embedding of entities according to the relationship connection of entities in the market dishonesty information knowledge graph includes: Step 1: Design a relation domain embedding and a relation range embedding for each basic relation; the relation domain embedding is used to transmit the hidden information of the basic relation when it is output by the head entity; the relation range embedding is used to transmit the hidden information of the basic relation when it is input by the tail entity; Step 2: For any entity, the relationship domain embedding and the relationship range embedding that reflect the input-output relationship are weighted averaged to obtain the type-level information embedding of the entity; the calculation formula is: ; in, Indicates e Type-level information embedding of entities; express For e The weight of an entity; express For e The weight of an entity; Indicates r The relational domain embedding of basic relations; Indicates r The relational scope of each basic relation is embedded; Indicates e The total number of basic relations output by each entity; Indicates e The total number of basic relations entered for each entity.

3. The method for embedding market dishonesty information knowledge graph inductive reasoning according to claim 2 is characterized in that: The step of constructing the instance-level information embedding of the entity according to the type-level information embedding of the entity and the neighborhood information of the corresponding entity comprises: Step 1: The neighborhood information of the entity includes the structural information of the multi-hop neighbors of the entity. The structural information and the type-level information of the entity are embedded and input into the graph neural network to obtain the hidden entity representation of the entity. l The update formula of the layer is: ; in, Indicates e The entity in l Hidden entity representation of the layer; ReLU activation function. Indicates all e The number of neighbor triplets that an entity is directly connected to; Indicates l Layer r The transformation matrix of the basic relations; Indicates l The self-circular transformation matrix of entities in the layer; Indicates e The entity in l- The hidden entity representation of the first layer; the type-level information of the entity is embedded as e The hidden entity representation of entities at layer 0; h Representation and Entity e Directly connected entities , Representing Entities h In the l- Hidden entity representation at layer 1; Step 2: Use the skip knowledge matrix to concatenate with the hidden entity representation to obtain the instance-level information embedding of the entity; the skip knowledge matrix is ​​used to convert the concatenated hidden entity representation into an entity embedding; the calculation formula is: ; in, Indicates e Instance-level information embedding of entities; represents the jump knowledge matrix; Represents vector concatenation.

4. The method for embedding market dishonesty information knowledge graph inductive reasoning according to claim 1 is characterized in that: The meta-learning based on the market dishonesty information knowledge graph, the comprehensive relationship embedding, and the instance-level information embedding of the entity to obtain the knowledge graph embedding model includes: Step 1: Use a random walk method to sample a group of sub-knowledge graphs from the market dishonesty information knowledge graph, and re-label the entities in each sub-knowledge graph as invisible entities; Step 2: Split the hyperrelation triples in any processed sub-knowledge graph into query triples and support triples to construct a meta-learning task; the expression of the meta-learning task is: ; ; in, Indicates i a meta-learning task; Represents the set of entities in the market dishonesty information knowledge graph; Represents the set of relations in the knowledge graph of market dishonesty information; Indicates i Support triplets for each meta-learning task; Indicates i query triples for a meta-learning task; s Represents the source knowledge graph; Represents the sub-knowledge graph sampled by the meta-task from the source knowledge graph; Represents the entity set of the source knowledge graph; Represents the relation set of the source knowledge graph; Step 3: In each meta-learning task, an entity embedding is obtained according to the support triples, and a loss function is calculated according to the query triples, the comprehensive relation embedding, and the instance-level information embedding of the entity; Step 4: constructing a meta-training objective of the meta-learning task based on the entity embedding and the loss function; Step 5: Train the meta-learning task based on the meta-training objective to obtain the knowledge graph embedding model.

5. The method for embedding market dishonesty information knowledge graph inductive reasoning according to claim 4 is characterized in that: The expression of the loss function is: ; ; ; in, represents the loss function; represents a positive triple; h represents the head entity of the positive triple; represents the relationship of fusion auxiliary information; t represents the tail entity of the triplet; Indicates i negative triples; Indicates i The head entity of a negative triple; Indicates i The tail entity of a negative triple; Indicates i query triples for a meta-learning task; Represents the sigmoid function; represents the marginal bias; represents the positive triple scoring function; represents the negative triple scoring function; Represents the self-adversarial weight value of the negative triple; represents the temperature coefficient of negative triplet sampling; Represents instance-level information embedding of the head entity; represents comprehensive relational embedding; Represents instance-level information embedding of the tail entity; k Represents the number of negative triplets.

6. The method for embedding market dishonesty information knowledge graph inductive reasoning according to claim 5 is characterized in that: The expression of the meta-training objective is: ; ; ; ; in, represents the first learning parameter; represents the second learning parameter; represents the third learning parameter; represents the fourth learning parameter; represents the comprehensive relation embedding matrix in the current task; m represents the number of meta-learning tasks; Indicates i The training loss for each meta-learning task; Represents the generation process of embedding instance-level information in query triples And the generation process of comprehensive relational embedding Generation losses; Indicates i Contrastive learning loss in a meta-learning task; Indicates i The loss function in each meta-learning task; n express The number of positive samples; Represents any relation embedding that incorporates auxiliary information; Indicates j positive samples; Indicates k negative samples; N represents the number of all super-relation triples in the batch; Indicates cosine similarity calculation; represents the temperature function.

Citation Information

Patent Citations

  • Common disease knowledge graph link prediction method fusing entity description

    CN117371525A

  • Knowledge graph completion method and related equipment

    CN118278519A