CN114579756A - Link prediction method based on reinforced entity structure feature representation learning - Google Patents

Link prediction method based on reinforced entity structure feature representation learning Download PDF

Info

Publication number
CN114579756A
CN114579756A CN202210137394.XA CN202210137394A CN114579756A CN 114579756 A CN114579756 A CN 114579756A CN 202210137394 A CN202210137394 A CN 202210137394A CN 114579756 A CN114579756 A CN 114579756A
Authority
CN
China
Prior art keywords
entity
relation
relationship
vector
learning
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202210137394.XA
Other languages
Chinese (zh)
Inventor
胡文轩
马宗民
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Aeronautics and Astronautics
Original Assignee
Nanjing University of Aeronautics and Astronautics
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University of Aeronautics and Astronautics filed Critical Nanjing University of Aeronautics and Astronautics
Priority to CN202210137394.XA priority Critical patent/CN114579756A/en
Publication of CN114579756A publication Critical patent/CN114579756A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
    • G06F16/367—Ontology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/35—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Mathematical Physics (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Databases & Information Systems (AREA)
  • Animal Behavior & Ethology (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Machine Translation (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

The invention discloses a link prediction method based on reinforced entity structure feature representation learning, which comprises the steps that firstly, a system generates an initialization vector according to entity data of a knowledge map, then the initialization vector of an entity is input into an entity representation learning framework, in the representation learning framework, the initial vector of the entity extracts the relation feature of the entity through an RGCN model, then a GCN model and a GAT model are combined to extract and learn the entity relation structure feature reflected by the entity structure feature to supplement the relation feature information, and multiple gating mechanisms are adopted to perform multi-feature fusion; finally, inputting the entity vector into a link prediction model, generating a prediction relation entity triple and verifying a prediction result; the invention effectively solves the problems of difficult multi-source information fusion and insufficient relation structure characteristics faced by a link prediction system, and the relation characteristic information contained in the entity characteristic vector is enough to identify the entity corresponding to the complex relation, thereby improving the link prediction effect.

Description

Link prediction method based on reinforced entity structure feature representation learning
Technical Field
The invention relates to the technical field of link prediction, in particular to a link prediction method based on reinforced entity structural feature representation learning.
Background
The knowledge graph is a knowledge base which displays entities and relations between the entities in the real world in a graph structure mode, nodes in the knowledge graph represent the entities in the real world, and edges between the nodes represent the relations between the entities. Although knowledge graphs can provide high-quality structured data, most knowledge graphs have sparse graph structures due to various factors, and implicit relations among a large number of entities are not represented. Therefore, the link prediction is an effective method for completing the knowledge graph, perfecting the knowledge representation in the knowledge graph and improving the quality of the knowledge graph.
Link prediction refers to predicting the relationship which may exist between entities through the relationship entities in the existing knowledge graph. Much of the recent research has been around how to better represent the knowledge-graph to improve the link prediction. The link prediction method based on entity representation learning is generally divided into an encoder part and a decoder part: in the encoder part, entity data are input into an entity representation learning model for learning, learning is carried out according to the existing normal triples of a knowledge graph in the learning process, and an objective function is set so that when the generated entity representation vector is used for calculating a score function, scores of relation entity triples which are correctly predicted are as high as possible, scores of relation entity triples which are incorrectly predicted are as low as possible, and finally the entity representation vector is obtained through learning; in the decoder part, the entity representation is learned to obtain an entity representation vector, and the score of the prediction triple (head entity, relation, tail entity) is calculated through a score function and is used as a judgment basis for judging whether the triple is a correct triple.
The link prediction method based on entity representation learning is mainly classified into a method based on a TransE model and a method based on a graph convolution network at present. In the learning process of the link prediction method based on the TransE model, the relation between an entity and a relationship is mainly considered, and the entity is usually modeled and feature-learned according to a head entity vector + a relationship vector, namely a tail entity vector, and a corresponding deformation equation.
In the entity representation learning process, the link prediction method based on the graph convolution network models and learns the relationship between entities according to the feature representation of the adjacent node of the central entity, obtains an entity representation vector, and then calculates the score condition of the corresponding prediction triple through a score function to be used as the judgment basis for judging whether the prediction triple is the correct triple.
The current link prediction method based on entity representation learning generally has the following problems:
(1) complex relational modeling is difficult: in real atlas data, the types of relationships between entities are very complex and the types of relationships are also very complex. In the prior link prediction method based on entity representation learning, only the simplest one-to-one type of relation is generally considered, but the modeling capability of a large number of one-to-many, many-to-many and many-to-one type of relations existing in real map data is not enough.
(2) The fusion of multi-source characteristic information is difficult: learning of the characteristic information of an entity is often inadequate. Due to the fact that the distinguishing degree between the entity feature vectors is insufficient due to the fact that the feature information in the entity representation is insufficient, when link prediction is conducted on sparse map data, wrong relation entity triples are predicted among entities which should not have corresponding relations.
(3) Knowledge-graph sparseness results in fewer positive samples: in the real map data, entities and relations are quite sparse, and the number of different relations is large. This results in the lack of sufficient formal relationship entity triples as samples of learning in the course of entity representation learning. For the relations with less occurrence times in the relation triples, in the entity representation learning process, due to the lack of the training sample, the difference characteristics of different relations are difficult to extract, and the link prediction effect is reduced.
Therefore, in order to solve the above three problems, the new link prediction method based on entity representation must take corresponding consideration and propose corresponding countermeasures in the entity representation learning process.
Disclosure of Invention
The purpose of the invention is as follows: in view of the problems in the background art, the present invention provides a link prediction method based on reinforced entity structural feature representation learning, which is based on the excellent capability of graph-convolution network to the entity data learning features, firstly, the relationship features of the entity are extracted and learned, and then, the adjacent structural features of the entity are correspondingly processed as the supplement and enrichment of the entity relationship features, so as to improve the representation capability of the relationship features in the entity representation; and finally, putting the entity representation and the relation matrix into a score function to calculate the score of the predicted triple to be used as a basis for judging whether the triple is correct or not.
The technical scheme is as follows: in order to achieve the purpose, the invention adopts the technical scheme that:
a link prediction method based on reinforced entity structural feature representation learning comprises the following steps:
step S1, inputting the initialization vector generated by the knowledge graph entity data into an RGCN model, extracting entity relationship characteristics and learning, and acquiring an entity relationship characteristic expression vector;
step S2, supplementing the entity relation characteristic based on the entity adjacent structure characteristic;
respectively inputting initialization vectors generated by knowledge graph entity data into a GCN model and a GAT model; in each layer of the entity representation learning network, the GCN model extracts and learns the neighbor structure characteristics of the central entity node, the GAT model gives weight according to the importance degree of the entity node to reflect the importance degree of the adjacent node, and the characteristics of the entity adjacent node are extracted to reflect the importance degree condition of the corresponding relation according to the importance degree of the adjacent node; the importance degree characteristics obtained by the GAT model are used as supplement of neighbor structure characteristics obtained by the GCN model, the two characteristics are aggregated based on an average gating mechanism, and a relational structure characteristic expression vector of an entity is obtained;
step S3, aggregating the entity relationship characteristic expression vector obtained in step S1 and the relationship structure characteristic expression vector in step S2 again based on Res gate gating, and obtaining an entity expression vector; in the learning and training process, the effect of representing vectors by learning and training entities is improved by reinforcing the sampling of negative samples as supplement to the deficiency of positive samples;
step S4, performing link prediction based on the entity expression vector;
the scoring function based on the DistMult decomposition is selected to calculate scores for the predicted triplet of relational entities to determine whether it can be considered as the correct triplet.
Further, the method for specifically extracting the entity relationship features in step S1 includes:
firstly, aiming at adjacent nodes of a central entity node, classifying according to the relationship between the central node and the corresponding adjacent nodes; then setting a learnable relation conversion matrix aiming at the relation between the central node and the corresponding adjacent node; finally, adding a reflexive relation for keeping the self characteristics of the nodes in the forward propagation learning process to the central node, wherein the entity represents the learning forward propagation formula as follows:
Figure BDA0003505478830000031
wherein the content of the first and second substances,
Figure BDA0003505478830000032
an entity of a next layer node i in the network represents a vector; σ denotes the activation function, here relu (); r represents a set of all the relations of the map; n is a radical ofrRepresenting a set of neighbor nodes corresponding to the node i under the relation r; c. Ci,rDenoted is a regularization constant;
Figure BDA0003505478830000033
representing a relationship characteristic transformation parameter at the layer;
Figure BDA0003505478830000034
representing a feature representation of the neighboring node j; wlRepresenting a learnable parameter matrix;
Figure BDA0003505478830000035
the representation is the feature representation obtained by the node i at the previous layer.
Further, the average gating mechanism fusion feature information formula adopted in step S2 is expressed as follows:
hi＝(1-α)hi1+αhi2
wherein h isiIs the relationship structure characteristic information of the entity i; h isi1、hi2The relationship structure characteristic vectors of the entity i are obtained by GCN and GAT models respectively; α is a fusion coefficient in the range of (0, 1)]And the method is used for aggregation of the relationship distribution characteristics and the relationship importance degree characteristics.
Further, the Res gate gating mechanism adopted in step S3 fuses the feature information formula as follows:
hi＝W2(σ(W1hi1+b))+hi2
wherein h isiIs the final feature representation vector of entity i; h isi1、hi2Respectively representing a relation structure characteristic vector and a relation characteristic vector model of the entity i; w1、W2Is a learnable parameter matrix; σ is an activation function, and relu () is selected as the activation function.
Further, the score function based on the DistMult decomposition is expressed as follows:
Figure BDA0003505478830000041
wherein h ishA feature representation vector representing a head entity h in the triplet; h istA feature representation vector representing a tail entity t in the triplet; rrRepresenting a diagonal matrix corresponding to the relation r, namely a relation conversion matrix; by using the DistMult decomposition, the head-to-tail entity feature vectors and the relation r are projected to a specific value as a probability score for the correctness of the link prediction triplet.
Has the advantages that:
the link prediction method for reinforcing entity structural feature representation learning provided by the invention creatively utilizes entity adjacent structural features to supplement entity relationship features, learns entity representation on the basis, and fuses feature information through a multi-gating mechanism, so that the relationship feature information in the entity representation vector obtained by learning is sufficient, and the entity vector corresponding to the complex relationship has high representation precision. When the score function is calculated through link prediction, the relation feature information contained in the entity feature vector is enough to identify the entity corresponding to the complex relation, and the link prediction effect is obviously improved.
Drawings
FIG. 1 is a flow chart of a link prediction method based on reinforced entity structural feature representation learning provided by the present invention;
FIG. 2 is a flow chart of a method for extracting and learning structural features of relationships according to the present invention.
Detailed Description
The present invention will be further described with reference to the accompanying drawings. It is to be understood that the embodiments described are only a few embodiments of the present invention, and not all embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
The method supplements the entity relationship characteristics by using the entity adjacent structure characteristics, performs representation learning on entity data, and uses the entity representation obtained by learning as a triple score calculation basis of link prediction, wherein the main flow is shown in fig. 1. The method provided by the invention is mainly divided into an encoding part and a decoding part. In the encoding process of the entity, the entity representation learning is carried out by inputting the initialized representation of the entity into an improved RGCN model and a structural feature learning network consisting of a GAT model and a GCN model respectively: the entity relation characteristic representation is obtained through the RGCN model, and the entity structure characteristic information obtained in the structure characteristic learning network can extract the relation condition of the entity reflected in the structure characteristic to obtain the entity relation structure characteristic. In the learning process of the relationship characteristics and the relationship structure characteristics, a sampling method for reinforcing negative example sample collection is used for improving the learning effect of positive example relationship entity triples, finally, the learned multi-source characteristic representation realizes information fusion through a special gating mechanism, and the obtained entity representation can sufficiently distinguish the complex relationships among the entities. In a decoding part, the invention uses a score function based on DistMult decomposition to score the predicted relation entity triples to obtain the possibility score of the predicted triples being correct triples, and the higher the score is, the more probable the predicted triples are considered to be correct. The following table 1 shows a specific flow of the link prediction method provided by the present invention:
TABLE 1 Link prediction method flow sheet
Figure BDA0003505478830000051
An example is provided below to further describe the link prediction method of the present invention:
and step S1, inputting the initialization vector generated by the knowledge graph entity data into an RGCN model, extracting entity relationship characteristics, learning and acquiring an entity relationship characteristic expression vector.
Firstly, the adjacent nodes of the central entity node are classified according to the relationship between the central node and the adjacent nodes. A learnable relationship transformation matrix is then set for the relationships contained between the central entity node and the neighboring nodes. Finally, adding a reflexive relation for keeping the self characteristics of the nodes in the forward propagation learning process to the central node, wherein the entity represents the learning forward propagation formula as follows:
Figure BDA0003505478830000052
wherein the content of the first and second substances,
Figure BDA0003505478830000053
the entities of the next level node i in the network represent vectors. σ denotes the activation function, here relu (). R represents a set of all relationships of the map. N is a radical ofrAnd representing the set of neighbor nodes corresponding to the node i under the relation r. c. Ci,rA regularization constant is indicated.
Figure BDA0003505478830000061
The relational feature transformation parameters at that level are represented.
Figure BDA0003505478830000062
A representation of the characteristics of the neighboring node j is shown. WlRepresenting a learnable parameter matrix.
Figure BDA0003505478830000063
The representation is the feature representation obtained by the node i at the previous layer.
And step S2, supplementing the entity relation characteristic based on the entity adjacent structure characteristic.
And respectively inputting the initialization vectors generated by the knowledge graph entity data into the GCN model and the GAT model. In each layer of the entity representation learning network, the GCN model extracts and learns the neighbor structure characteristics of the central entity node, the GAT model gives weight according to the importance degree of the entity node to reflect the importance degree of the adjacent node, and the characteristics of the entity adjacent node are extracted to reflect the importance degree condition of the corresponding relation according to the importance degree of the adjacent node. And taking the importance degree characteristics obtained by the GAT model as the supplement of the neighbor structure characteristics obtained by the GCN model, aggregating the two characteristics based on an average gating mechanism, and obtaining the relation structure characteristic expression vector of the entity.
In this step, the neighbor structure features of the central node are first learned and extracted using the underlying GCN model. Meanwhile, since the respective degrees of importance of the adjacent nodes to the central node are different, the degrees of importance of the adjacent nodes can also be used to represent the degrees of importance of the relationship. Therefore, the present embodiment uses the GAT model at the same time, and captures the structural features according to the difference in importance of the neighboring nodes, so as to reflect the difference in importance between the relationships owned by the central entity. And then, an entity representation obtained by GAT learning is used as a supplement of the structural characteristics obtained by the GCN model, and the entity representation and the GCN model are fused according to a certain weighting by a gating mechanism, so that the neighbor structural characteristic information can reflect the distribution and quantity condition characteristics of the overall relationship and can reflect the difference of the importance degrees of corresponding relationships of different nodes. As shown in fig. 2, in the learning module of the relationship structure feature information, the present invention performs corresponding improvement and adjustment on the GCN model and the GAT model, and learns the adjacent structure feature and the relationship importance degree feature that reflect the central entity relationship distribution feature in a parallel manner in the whole structure feature learning framework, and then performs aggregation of the relationship distribution feature and the relationship importance degree feature through average gating.
The average gating mechanism fusion feature information formula is expressed as follows:
hi＝(1-α)hi1+αhi2
wherein h isiIs the relationship structure characteristic information of the entity i. h isi1、hi2The relationship structure feature vectors of the entity i are obtained by GCN and GAT models respectively. α is a fusion coefficient in the range of (0, 1)]And the method is used for aggregation of the relationship distribution characteristics and the relationship importance degree characteristics. Because the relationship structure feature information of the central entity is related to the adjacent nodes, the finally obtained relationship structure feature expression vector needs to contain enough adjacent structure feature information, and meanwhile, because of the diversity of the relationship, the adjacent nodes need to be learned according to the importance, and the importance of the corresponding relationship is reflected by the importance of the adjacent nodes, so that the difference between different relationships can be more obvious. By using average gating, the features obtained by the GAT model according to the learning of the importance of the adjacent nodes can take the importance degree of the adjacent nodes as the importance degree of each relationship corresponding to the central entity, so that differential feature information on the relationship structure can be embodied. Meanwhile, adjacent node features which are ignored due to low attention coefficients in the GAT model can be retained in the node feature representation obtained by GCN learning, so that the loss of feature information in the fusion process is reduced as much as possible.
And step S3, aggregating the entity relationship characteristic representation vector obtained in step S1 and the relationship structure characteristic representation vector in step S2 again based on Res gate gating, and obtaining the entity representation vector. And in the learning training process, the effect of the expression vector of the learning training entity is improved by an improved negative sampling method.
The adopted Res gate control mechanism fusion characteristic information formula is expressed as follows:
hi＝W2(σ(W1hi1+b))+hi2
wherein h isiIs the final feature representation vector for entity i. h isi1、hi2Respectively, a relational structure feature vector and a relational feature vector model of the entity i. W1、W2Is a learnable parameter matrix. σ is an activation function, and relu () is selected as the activation function. Because the relationship features are most important for link prediction, the relationship features should be retained as much as possible in the fusion process, and in order that the relationship structure feature information can be supplemented as the relationship features and the entity relationship features contained in the whole vector are not weakened due to the introduction of the relationship structure features, the adjusted Res gate control is used as a fusion mode of the relationship feature information and the relationship structure feature information.
In the process of sampling data when the model is trained, a new negative sampling method is used. Because the number of correct relationship entity triples as positive examples is small due to the sparsity of the real map data, the link prediction method provided in this embodiment is used as a supplement to the deficiency of the positive example samples by enhancing the sampling of the negative example samples during learning. In the previous research, most negative example sampling methods replace the head entity or the tail entity in each correct relationship triplet randomly with other entities, so that the original correct relationship triplet becomes an incorrect triplet sample. The negative sampling method used in this embodiment randomly replaces the relationship while replacing the head and tail entities, so that a correct relationship entity triplet cannot be generated when a negative triplet is generated by random replacement as much as possible.
Step S4, link prediction is performed based on the entity expression vector.
After the entity feature representation vector is obtained, the entity feature representation vector can be decoded, and link prediction is completed. The embodiment chooses to score the predicted relationship entity triplet calculation based on the scoring function of the DistMult decomposition to determine whether it can be regarded as the correct triplet.
The score function based on DistMult decomposition is represented as follows:
Figure BDA0003505478830000081
wherein h ishThe feature representing the head entity h in the triplet represents the vector. h istThe feature representing the tail entity t in the triplet represents the vector. RrAnd (3) representing a diagonal matrix corresponding to the relation r, and setting the diagonal matrix as a learnable parameter matrix, namely a relation conversion matrix. By using the DistMult decomposition, the head-to-tail entity feature vectors and the relation r are projected to a specific value as a probability score for the correctness of the link prediction triplet.
The above description is only of the preferred embodiments of the present invention, and it should be noted that: it will be apparent to those skilled in the art that various modifications and adaptations can be made without departing from the principles of the invention and these are intended to be within the scope of the invention.

Claims (5)

1. A link prediction method based on reinforced entity structural feature representation learning is characterized by comprising the following steps:
step S1, inputting the initialization vector generated by the knowledge graph entity data into an RGCN model, extracting entity relationship characteristics and learning, and acquiring an entity relationship characteristic expression vector;
step S2, supplementing the entity relation characteristic based on the entity adjacent structure characteristic;
respectively inputting initialization vectors generated by knowledge graph entity data into a GCN model and a GAT model; in each layer of the entity representation learning network, the GCN model extracts and learns the neighbor structure characteristics of the central entity node, the GAT model gives weight according to the importance degree of the entity node to reflect the importance degree of the adjacent node, and the characteristics of the entity adjacent node are extracted to reflect the importance degree condition of the corresponding relation according to the importance degree of the adjacent node; the importance degree characteristics obtained by the GAT model are used as supplement of neighbor structure characteristics obtained by the GCN model, the two characteristics are aggregated based on an average gating mechanism, and a relational structure characteristic expression vector of an entity is obtained;
step S3, aggregating the entity relationship characteristic expression vector obtained in step S1 and the relationship structure characteristic expression vector in step S2 again based on Res gate gating to obtain an entity expression vector; in the learning and training process, the effect of the representation vector of the learning and training entity is improved by reinforcing the sampling of the negative sample as the supplement of the shortage of the positive sample;
step S4, performing link prediction based on the entity expression vector;
the scoring function based on the DistMult decomposition is selected to calculate scores for the predicted triplet of relational entities to determine whether it can be considered as the correct triplet.
2. The method for predicting links based on reinforced entity structural feature representation learning as claimed in claim 1, wherein the method for specifically extracting entity relationship features in step S1 includes:
firstly, classifying adjacent nodes of a central entity node according to the relationship between the central node and the corresponding adjacent nodes; then, setting a learnable relation conversion matrix aiming at the relation between the central node and the corresponding adjacent node; finally, adding a reflexive relation for keeping the self characteristics of the nodes in the forward propagation learning process to the central node, wherein the entity expression learning forward propagation formula is as follows:
Figure FDA0003505478820000011
wherein, the first and the second end of the pipe are connected with each other,
Figure FDA0003505478820000012
an entity of a next layer node i in the network represents a vector; σ represents an activation function; r represents a set of all the relations of the map; n is a radical ofrRepresenting a set of neighbor nodes corresponding to the node i under the relation r; c. Ci,rDenoted is a regularization constant;
Figure FDA0003505478820000013
representing a relationship characteristic transformation parameter at the layer;
Figure FDA0003505478820000014
representing a characteristic representation of the adjacent node j; wlRepresenting a learnable parameter matrix;
Figure FDA0003505478820000021
the representation is the feature representation obtained by the node i at the previous layer.
3. The method for predicting links based on reinforced entity structural feature representation learning of claim 1, wherein the mean gating mechanism adopted in step S2 is fused with feature information formula as follows:
hi＝(1-α)hi1+αhi2
wherein h isiIs the relationship structure characteristic information of the entity i; h isi1、hi2The relationship structure characteristic vectors of the entity i are obtained by GCN and GAT models respectively; α is a fusion coefficient in the range of (0, 1)]The method is used for aggregation of the relationship distribution characteristics and the relationship importance degree characteristics.
4. The method for predicting links based on reinforced entity structural feature representation learning of claim 1, wherein the Res gate gating mechanism adopted in step S3 fuses the feature information formula as follows:
hi＝W2(σ(W1hi1+b))+hi2
wherein h isiIs the final feature representation vector of entity i; h isi1、hi2Respectively a relation structure feature vector and a relation feature vector model of the entity i; w1、W2Is a learnable parameter matrix; σ is an activation function, and relu () is selected as the activation function.
5. The method according to claim 1, wherein the score function based on DistMult decomposition is expressed as follows:
Figure FDA0003505478820000022
wherein h ishA feature representation vector representing a head entity h in the triplet; h istA feature representation vector representing a tail entity t in the triplet; rrRepresenting a diagonal matrix corresponding to the relation r, namely a relation conversion matrix; by using the DistMult decomposition, the head-to-tail entity feature vectors and the relation r are projected to a specific value as a probability score for the correctness of the link prediction triplet.
CN202210137394.XA 2022-02-15 2022-02-15 Link prediction method based on reinforced entity structure feature representation learning Pending CN114579756A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210137394.XA CN114579756A (en) 2022-02-15 2022-02-15 Link prediction method based on reinforced entity structure feature representation learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210137394.XA CN114579756A (en) 2022-02-15 2022-02-15 Link prediction method based on reinforced entity structure feature representation learning

Publications (1)

Publication Number Publication Date
CN114579756A true CN114579756A (en) 2022-06-03

Family

ID=81775382

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210137394.XA Pending CN114579756A (en) 2022-02-15 2022-02-15 Link prediction method based on reinforced entity structure feature representation learning

Country Status (1)

Country Link
CN (1) CN114579756A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115391563A (en) * 2022-09-01 2022-11-25 广东工业大学 Knowledge graph link prediction method based on multi-source heterogeneous data fusion

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115391563A (en) * 2022-09-01 2022-11-25 广东工业大学 Knowledge graph link prediction method based on multi-source heterogeneous data fusion
CN115391563B (en) * 2022-09-01 2024-02-06 广东工业大学 Knowledge graph link prediction method based on multi-source heterogeneous data fusion

Similar Documents

Publication Publication Date Title
CN106650789B (en) Image description generation method based on depth LSTM network
CN111738942A (en) Generation countermeasure network image defogging method fusing feature pyramid
CN110321361B (en) Test question recommendation and judgment method based on improved LSTM neural network model
CN111126488A (en) Image identification method based on double attention
CN108711141A (en) The motion blur image blind restoration method of network is fought using improved production
CN104318214A (en) Cross view angle face recognition method based on structuralized dictionary domain transfer
CN104463101A (en) Answer recognition method and system for textual test question
CN112464004A (en) Multi-view depth generation image clustering method
CN112766315B (en) Method and system for testing robustness of artificial intelligence model
CN112784929B (en) Small sample image classification method and device based on double-element group expansion
CN111949764B (en) Knowledge graph completion method based on bidirectional attention mechanism
CN106530341A (en) Point registration algorithm capable of keeping local topology invariance
CN113190688A (en) Complex network link prediction method and system based on logical reasoning and graph convolution
CN112989942A (en) Target instance segmentation method based on traffic monitoring video
CN111680702A (en) Method for realizing weak supervision image significance detection by using detection frame
CN109214444B (en) Game anti-addiction determination system and method based on twin neural network and GMM
CN115620010A (en) Semantic segmentation method for RGB-T bimodal feature fusion
CN115330620A (en) Image defogging method based on cyclic generation countermeasure network
CN114579756A (en) Link prediction method based on reinforced entity structure feature representation learning
CN114998602A (en) Domain adaptive learning method and system based on low confidence sample contrast loss
CN111339320A (en) Knowledge graph embedding and reasoning method introducing entity type automatic representation
CN114819143A (en) Model compression method suitable for communication network field maintenance
CN114821299B (en) Remote sensing image change detection method
CN106529604A (en) Adaptive image tag robust prediction method and system
CN103281555B (en) Half reference assessment-based quality of experience (QoE) objective assessment method for video streaming service

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination