CN110704636A - Improved Node2 vec-based knowledge graph vector representation method - Google Patents
Improved Node2 vec-based knowledge graph vector representation method Download PDFInfo
- Publication number
- CN110704636A CN110704636A CN201910926716.7A CN201910926716A CN110704636A CN 110704636 A CN110704636 A CN 110704636A CN 201910926716 A CN201910926716 A CN 201910926716A CN 110704636 A CN110704636 A CN 110704636A
- Authority
- CN
- China
- Prior art keywords
- node
- entity
- nodes
- knowledge graph
- relational
- 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.)
- Granted
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
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/28—Databases characterised by their database models, e.g. relational or object models
- G06F16/284—Relational databases
- G06F16/288—Entity relationship models
-
- 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
Abstract
The invention discloses an improved Node2 vec-based knowledge graph vector representation method, which comprises the following steps: step one, reconstructing a knowledge graph G' ═ E, R and L; step two, setting transition probability; step three, setting a wandering path; setting a training model according to the type of the node and the position of the node; fifthly, optimizing training parameters; and step six, outputting vectorization representation of the nodes through multiple times of training. The invention has the beneficial effects that: and repeatedly training by combining the reconstructed knowledge graph structure, and fully training the wandering sequence to obtain the semantic information of the nodes according to a mode of cross-wandering the entity and the relation, so that the property significance of the network topology structure can be obtained while the network topology structure is obtained. The homogeneous structure network is changed into the heterogeneous structure network, the characteristics of the knowledge graph network structure are better adapted, the walking sequence is closer to the natural language structure, and the prediction result is more accurate.
Description
Technical Field
The invention relates to a knowledge graph vector representation method, in particular to an improved knowledge graph vector representation method based on a Node2 vec.
Background
The knowledge graph was proposed by Google corporation in 2012, and nodes in the knowledge graph are assigned meanings of entities, and edges between the entities represent relationships between the entities, as compared to a conventional graph structure. Knowledge-graphs can thus link together different kinds of information, providing the ability to analyze problems from a "relational" perspective.
In order to better process multi-relation data, a knowledge representation learning technology is introduced, namely representation learning is carried out on the relation between entities. The KG Embedding correlation algorithm also takes place. KG Embedding's general steps 1) represent entities and relationships in the graph; 2) defining a scoring function; 3) vector representations of entities and relationships are learned.
The TransE model is a classic algorithm for processing KG embedding, and since the time of Bordes et al 2013, a series of models are generated to improve and supplement the TransE model, such as TransH, TransG and the like.
In the traditional method for modeling the triple (head, relation, tail) in the training knowledge base, parameters are more, so that model interpretability is low, and the problem of overfitting is easy to occur in the training process. The TransE model improves the loss function, a reward and punishment mechanism is added, the prediction result is maximized in a staggered way as far as possible, and the defects that the training parameters are complex and difficult to expand in the traditional method are overcome. The TransH algorithm proposed by Zhen Wang et al adds two relational correlation matrices on the basis of TransE for representing the head and the tail, does not map the relations to another space, but represents the relations by vectors, and solves the problem of modeling of one-to-many, many-to-one, many-to-many relations by TransE. Guolang Ji et al improved and extended the TransE model on the link prediction problem, and proposed the TransD model, which defines a mapping matrix for each relationship, and increased the computational complexity while improving the prediction accuracy. Hao Xian et al propose a TransA algorithm, redefine the loss metric function, and solve the problem of dimensional noise by adding a matrix Wr and using an ellipse equivalent hyperplane. Shizhu He et al propose a KG2E algorithm, which uses Gaussian distribution to represent entities and relations, and uses covariance of the Gaussian distribution to represent uncertainty between the entities and the relations, thereby improving accuracy of link prediction.
Research shows that the method has certain limitations, has some fundamental problems and is not solved. The first problem is that a representation vector with sufficient semantics cannot be obtained through training, and the second problem is that the complexity of the model and the accuracy of the model cannot be balanced.
Disclosure of Invention
The invention aims to solve the problems that the prior knowledge graph vector representation method comprises the following steps: the Node2 vec-based knowledge graph vector representation method has the advantages that sufficient representation vectors cannot be obtained through training, and the complexity of the model and the accuracy of the model cannot be balanced.
In a traditional knowledge graph network architecture, entities, semantic classes, content, attributes, and relationships are included. The entities, semantic classes, content, and attribute values form nodes in the knowledge-graph network graph, and the relationships are represented by edges linking the nodes. Knowledge-graphs are typically represented in triplets.
The prior knowledge-graph G ═ (E, R, S), where E represents the set of E different entities in the knowledge-graph, E ═ { E1, E2, E3 … …, E | }, the entities represented by nodes in the knowledge-graph network graph. R represents a set of relationships in the knowledgegraph, R ═ { R1, R2, R3 … …, R | }, in the network graph of the knowledgegraph, the relationships are usually represented by edges in the network graph, and one edge in the graph links two nodes, so as to form a triplet of "entity-relationship-entity".Representing triples in the knowledge-graph. In order to homogenize the knowledge-graph, the knowledge-graph network graph needs to be reconstructed.
The invention provides an improved Node2 vec-based knowledge graph vector representation method, which comprises the following steps:
step one, reconstructing a knowledge graph G ═ E, R, L, (E, R, L), where E is a set of entity type nodes in an original knowledge graph network, R is a relationship type node extracted from a set of relationships in an original structure, and L is an edge linking the entity node and the relationship node or the entity node and the entity node, and there is no practical significance, and the definition of the reconstructed knowledge graph network structure is as follows:
definition 1: for the reconstructed knowledge graph G, two types of nodes exist, namely a relational node and an entity node, no link exists between the relational nodes, the link between the entity node and the relational node is the most common link, and the link exists between the two entity nodes;
step two, setting transition probability: in order to apply a node2vec model to a knowledge graph, a knowledge graph network structure is improved in a homogenization mode, relational nodes are used for eliminating labels of edges, node types are changed from single types to two types of nodes, namely entities and relational types, a migration strategy needs to be improved, the migration strategy is different for different types of nodes, compared with a common social network, the knowledge graph network has certain characteristics, most of the nodes exist in an entity-relation-entity sequence, and an assumption 1 is given according to the rule, wherein the assumption is specifically as follows:
assume that 1: the node context in the knowledge graph has strong regularity, most of the nearest context nodes are relational nodes for entity nodes, and the nearest context nodes can only be entity type nodes for relational nodes;
different node migration strategies are given according to different node types;
and (3) entity node: if the current node is a solid type node S1Is through a relational node R2Obtained by jumping, i.e. solid type S1The last one of the nodes being a relational node R2,S1There are three cases of the next transition, and the transition probabilities are as follows:
1) jump back to the previous relational node R2This condition is of no significance to the knowledge-graph, so the transition probabilitiesIs 0;
2) from S1Jump to another relational node R associated therewith1This is desirable, and follows the assumption 1, so the hop probability is 1;
3) jump to the entity node S associated therewith2This is present but not common for this case, so the probability is set to 1/q;
the jump parameter α for the solid type node is as follows:
where α represents a transfer parameter, q is a training parameter, dtxRepresents the shortest path between node t and node x, and assumes a second order Markov condition is met, dtxCan only have one value between {0, 1, 2 };
relation node:
if the current node is a relational node, the α parameter is set as follows:
as can be seen from assumption 1, the relationship node can only be connected to the entity node, the current node is node R, and the previous node is node S1The next transition probability of the node R is three, specifically as follows:
1) jump to its previous node S1In this case, the probability is set to 0, which is not present in the knowledge map;
2) jump to and previous node S1Node S with an association2And the two are combined into a logical string of logical 'entity-relation-entity', and the transition probability is set to be 1;
3) jumping to another entity node S related to R3Forming a logic string of 'entity-relation' and 'relation-entity', and setting the logic probability to be 1/p;
the hop parameter α for a node type is as follows:
where α represents a transfer parameter, p is a training parameter, dtxRepresents the shortest path between node t and node x, and assumes a second order Markov condition is met, dtxCan only have one value between {0, 1, 2 };
step three, setting a wandering path: acquiring the transition probability of the relation node and the entity node through a formula, setting a walking path, and regarding the knowledge graph, taking a triple group of 'entity-relation-entity' as a construction basis of the walking path;
step four, setting a training model according to the type of the node and the position of the node: namely a CBOW model of a prediction relationship, a CBOW model of a prediction entity, a Skip-gram model of a prediction entity context and a Skip-gram model of a prediction relationship context;
step five, training parameter optimization: defining the influence of the node according to the degree of the node, enabling the node with larger influence to have more wandering times, enabling the wandering times to be correspondingly reduced for the nodes with smaller influence of the node, setting a threshold value, performing wandering training according to the maximum wandering times when the degree of the node is larger than the threshold value, reducing the wandering times according to the proportion of the influence if the degree of the node is smaller than the threshold value, and enabling the wandering times N of the node p to be larger than the threshold valuepThe definition is as follows:
wherein N ismaxIs the maximum number of walks, DpIs the degree of node p, DmaxThe node degree is the maximum degree in all nodes, and t is a set threshold value;
and step six, outputting vectorization representation of the nodes through multiple times of training.
The invention has the beneficial effects that:
the technical scheme provided by the invention adopts a Node2Vec method as a basic model, combines a reconstructed knowledge map structure to carry out repeated training, can fully train a wandering sequence to obtain semantic information of nodes according to a mode of crossing and wandering an entity and a relation, and can obtain the property significance while obtaining a network topological structure. The algorithm reduces the complexity while achieving the effects of the Trans series, does not need excessive mapping matrixes or transfer matrixes, directly learns through a semi-random walk strategy, and reduces the time complexity and the space complexity while ensuring the accuracy. The training model of KG2vec adopts CBOW and Skip-gram models to respectively carry out vectorization representation prediction on a solid node and a relational node. In order to improve the quality and speed of model training, a parameter of node degree is introduced to set the walking times. The nodes are divided into relational nodes and entity nodes, different migration and training modes are provided for different types of nodes, a homogeneous structure network is changed into a heterogeneous structure network, the characteristics of a knowledge graph network structure are better adapted, a migration sequence is closer to a natural language structure, and a prediction result is more accurate.
Drawings
FIG. 1 is a schematic diagram of the overall operation of the method of the present invention.
Fig. 2 is a schematic diagram of entity node hopping according to the present invention.
FIG. 3 is a diagram illustrating a relation node jumping situation according to the present invention.
FIG. 4 is a schematic diagram of a CBOW model of the predictive relationship of the present invention.
FIG. 5 is a schematic diagram of a CBOW model of a predicted entity according to the present invention.
FIG. 6 is a schematic diagram of a Skip-gram model for predicting entity context according to the present invention.
FIG. 7 is a schematic diagram of the Skip-gram model of the predictive relational context according to the present invention.
Detailed Description
Please refer to fig. 1 to 7:
in the WN18 dataset, the overall process of operation is performed:
step one, processing the data set, wherein WN18 comprises 18 relations and 40942 entity nodes, and in the original data, the entity nodes are stored in the form of node-relation-node triples. Now, the nodes in the data set are changed into entity nodes, and the relationship attributes in the data set are also used as nodes to form a new data set, for example, if there is a relationship of 3 between the node 1 and the node2, it will be indicated that there is a connection line between the node 1 and the node2 in the original knowledge graph, and the connection line has an attribute value of 3. For the modified knowledge graph, it can be represented that the entity node 1 is connected with the relational node 3, and the relational node 3 is also connected with the entity node 2. In the process of reconstructing the knowledge graph data set, the relationship and the nodes in the data set are changed into nodes, the original nodes are classified into entity node types, and the original relationship is classified into relationship node types. And transforming the original triple into a form of a binary group through a KB2edgelist.
And step two, carrying out experiments aiming at the p value and the q value in the walking strategy, selecting the optimal parameters and setting. For the data set, four values of 0.1, 1, 10 and 100 are selected as q values, under the condition that the value of p is determined, the influence of q on the recall rate is found to have no obvious regularity, and under different conditions, the recall rate has larger fluctuation, which shows that the influence of q on the effect of the random walk model is more random, and the initial value of q is finally set to be 100, because the recall rate has a rising trend along with the increase of the value of q in general; in addition, the optimized training strategy proposed herein does not play a significant role in the recall-q experiment, which laterally illustrates that in the random walk algorithm proposed herein, p has a greater impact on the model effect than q.
For the selection of the p value, four values of 0.1, 1, 10 and 100 are also selected, under the condition of determining the q value, the larger the initial value setting, the more the recall rate is increased, no matter the entity vector or the relation vector, which is completely consistent with the guess of the random game model, and meanwhile, we can also find that when p takes 10 and 100, the difference of the recall rate is very small, even under some conditions, the effect of taking 10 is better, so that the initial value of p is finally set as 10;
after p, q and the training parameters are set, parameter setting is respectively carried out on a node2vec model, a node2vec _ cbow _ improved model, a node2vec _ skip model and a node2vec _ skip _ improved model, and the training parameters are set to be the same form and numerical values in order to ensure the reliability of subsequent tests.
And step three, calculating the probability of skipping to the next node of each entity or relationship node after the p value and the q value are obtained. And (3) respectively using a formula 1 and a formula 2 to calculate the node jump probability, jumping all jumps in an entity-relationship-entity walking mode, finding the node with the maximum jump probability each time through a plurality of jumps, and forming a semi-random walking sequence. The current node being a solid type node S1Is through a relational node R2And (6) obtaining the jump. I.e. solid type S1The last one of the nodes being a relational node R2。S1There are three cases of the next transition, and the transition probabilities are as follows: 1) jump back to last relational node R2This case is not meaningful for the knowledge graph, so the transition probability is 0; 2) from S1Jump to another relational node R associated therewith1This situation is what we expect, and it is consistent with the assumption 1, so the hop probability is 1; 3) jump to the entity node S associated therewith2Since this is not common but present, the probability is set to 1/q. The current node is node R, the previous node is node S1. There are three types of next transition probabilities for node R: 1) jump to its previous node S1In this case, since the knowledge map does not exist, the probability is set to 0; 2) jump to and last node S1Node S with an association2And to be pieced together into a logical string of logical "entity-relationship-entity", which is expected to occur, so that the transition probability is set to 1; 3) jump to another entity node S related to R3Although this may occur when a logical string of "entity-relationship" and "relationship-entity" is formed, the logical probability is set to 1/p.
And step four, after the random walk sequence is obtained, training the random walk sequence through a CBOW model and a Skip-gram model respectively, and outputting the random walk sequence through an input layer, a hidden layer and softmax respectively, wherein a negative sampling method is adopted for training to obtain vectorization representation of all nodes. The trained node vectors are also classified into node type vectors and entity type vectors.
And step five, carrying out an accuracy and recall rate verification experiment, and respectively comparing the Node2Vec model, the TransE model and the recall rate value of the patent model method.
And step six, the recall rate result of the method is optimal no matter the entity node or the relation node.
Claims (1)
1. An improved Node2 vec-based knowledge graph vector representation method is characterized in that: the method comprises the following steps:
step one, reconstructing a knowledge graph G ═ E, R, L, (E, R, L), where E is a set of entity type nodes in an original knowledge graph network, R is a relationship type node extracted from a set of relationships in an original structure, and L is an edge linking the entity node and the relationship node or the entity node and the entity node, and there is no practical significance, and the definition of the reconstructed knowledge graph network structure is as follows:
definition 1: for the reconstructed knowledge graph G, two types of nodes exist, namely a relational node and an entity node, no link exists between the relational nodes, the link between the entity node and the relational node is the most common link, and the link exists between the two entity nodes;
step two, setting transition probability: in order to apply a node2vec model to a knowledge graph, a knowledge graph network structure is improved in a homogenization mode, relational nodes are used for eliminating labels of edges, node types are changed from single types to two types of nodes, namely entities and relational types, a migration strategy needs to be improved, the migration strategy is different for different types of nodes, compared with a common social network, the knowledge graph network has certain characteristics, most of the nodes exist in an entity-relation-entity sequence, and an assumption 1 is given according to the rule, wherein the assumption is specifically as follows:
assume that 1: the node context in the knowledge graph has strong regularity, most of the nearest context nodes are relational nodes for entity nodes, and the nearest context nodes can only be entity type nodes for relational nodes;
different node migration strategies are given according to different node types;
and (3) entity node: if the current node is a solid type node S1Is through a relational node R2Obtained by jumping, i.e. solid type S1The last one of the nodes being a relational node R2,S1There are three cases of the next transition, and the transition probabilities are as follows:
1) jump back to the previous relational node R2This case is not meaningful for the knowledge graph, so the transition probability is 0;
2) from S1Jump to another relational node R associated therewith1This is desirable, and follows the assumption 1, so the hop probability is 1;
3) jump to the entity node S associated therewith2This is present but not common for this case, so the probability is set to 1/q;
the jump parameter α for the solid type node is as follows:
where α represents a transfer parameter, q is a training parameter, dtxRepresents the shortest path between node t and node x, and assumes a second order Markov condition is met, dtxCan only have one value between {0, 1, 2 };
relation node:
if the current node is a relational node, the α parameter is set as follows:
as can be seen from assumption 1, the relationship node can only be connected to the entity node, the current node is node R, and the previous node is node S1The next transition probability of the node R is three, specifically as follows:
1) jump to its previous node S1For this caseThat is, there is no knowledge map, and the probability is set to 0;
2) jump to and previous node S1Node S with an association2And the two are combined into a logical string of logical 'entity-relation-entity', and the transition probability is set to be 1;
3) jumping to another entity node S related to R3Forming a logic string of 'entity-relation' and 'relation-entity', and setting the logic probability to be 1/p;
the hop parameter α for a node type is as follows:
where α represents a transfer parameter, p is a training parameter, dtxRepresents the shortest path between node t and node x, and assumes a second order Markov condition is met, dtxCan only have one value between {0, 1, 2 };
step three, setting a wandering path: acquiring the transition probability of the relation node and the entity node through a formula, setting a walking path, and regarding the knowledge graph, taking a triple group of 'entity-relation-entity' as a construction basis of the walking path;
step four, setting a training model according to the type of the node and the position of the node: namely a CBOW model of a prediction relationship, a CBOW model of a prediction entity, a Skip-gram model of a prediction entity context and a Skip-gram model of a prediction relationship context;
step five, training parameter optimization: defining the influence of the node according to the degree of the node, enabling the node with larger influence to have more wandering times, enabling the wandering times to be correspondingly reduced for the nodes with smaller influence of the node, setting a threshold value, performing wandering training according to the maximum wandering times when the degree of the node is larger than the threshold value, reducing the wandering times according to the proportion of the influence if the degree of the node is smaller than the threshold value, and enabling the wandering times N of the node p to be larger than the threshold valuepThe definition is as follows:
wherein N ismaxIs the maximum number of walks, DpIs the degree of node p, DmaxThe node degree is the maximum degree in all nodes, and t is a set threshold value;
and step six, outputting vectorization representation of the nodes through multiple times of training.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910926716.7A CN110704636B (en) | 2019-09-27 | 2019-09-27 | Improved Node2 vec-based knowledge graph vector representation method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910926716.7A CN110704636B (en) | 2019-09-27 | 2019-09-27 | Improved Node2 vec-based knowledge graph vector representation method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN110704636A true CN110704636A (en) | 2020-01-17 |
CN110704636B CN110704636B (en) | 2021-01-05 |
Family
ID=69196858
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910926716.7A Active CN110704636B (en) | 2019-09-27 | 2019-09-27 | Improved Node2 vec-based knowledge graph vector representation method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110704636B (en) |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111708881A (en) * | 2020-05-22 | 2020-09-25 | 国网天津市电力公司 | Text representation learning method introducing incidence relation |
CN112182249A (en) * | 2020-10-23 | 2021-01-05 | 四川大学 | Automatic classification method and device for aviation safety report |
CN112256801A (en) * | 2020-10-10 | 2021-01-22 | 深圳力维智联技术有限公司 | Method, system and storage medium for extracting key entities in entity relationship graph |
CN112463989A (en) * | 2020-12-11 | 2021-03-09 | 交控科技股份有限公司 | Knowledge graph-based information acquisition method and system |
CN113495963A (en) * | 2020-03-19 | 2021-10-12 | 复旦大学 | Embedded representation method and device of network security knowledge graph |
CN114283947A (en) * | 2021-12-27 | 2022-04-05 | 北京和兴创联健康科技有限公司 | Health management method and system suitable for surgical patients |
CN115617694A (en) * | 2022-11-30 | 2023-01-17 | 中南大学 | Software defect prediction method, system, device and medium based on information fusion |
CN116524723A (en) * | 2023-06-27 | 2023-08-01 | 天津大学 | Truck track anomaly identification method and system |
US20230306321A1 (en) * | 2022-03-24 | 2023-09-28 | Chengdu Qinchuan Iot Technology Co., Ltd. | Systems and methods for managing public place in smart city |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150379414A1 (en) * | 2014-06-27 | 2015-12-31 | Nuance Communications, Inc. | Utilizing large-scale knowledge graphs to support inference at scale and explanation generation |
CN107391542A (en) * | 2017-05-16 | 2017-11-24 | 浙江工业大学 | A kind of open source software community expert recommendation method based on document knowledge collection of illustrative plates |
CN109543034A (en) * | 2018-11-07 | 2019-03-29 | 中山大学 | Text Clustering Method, device and the readable storage medium storing program for executing of knowledge based map |
CN109933785A (en) * | 2019-02-03 | 2019-06-25 | 北京百度网讯科技有限公司 | Method, apparatus, equipment and medium for entity associated |
CN110162638A (en) * | 2019-04-12 | 2019-08-23 | 淮阴工学院 | A kind of expert's combination proposed algorithm based on figure vector |
CN110245204A (en) * | 2019-06-12 | 2019-09-17 | 桂林电子科技大学 | A kind of intelligent recommendation method based on positioning and knowledge mapping |
-
2019
- 2019-09-27 CN CN201910926716.7A patent/CN110704636B/en active Active
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20150379414A1 (en) * | 2014-06-27 | 2015-12-31 | Nuance Communications, Inc. | Utilizing large-scale knowledge graphs to support inference at scale and explanation generation |
CN107391542A (en) * | 2017-05-16 | 2017-11-24 | 浙江工业大学 | A kind of open source software community expert recommendation method based on document knowledge collection of illustrative plates |
CN109543034A (en) * | 2018-11-07 | 2019-03-29 | 中山大学 | Text Clustering Method, device and the readable storage medium storing program for executing of knowledge based map |
CN109933785A (en) * | 2019-02-03 | 2019-06-25 | 北京百度网讯科技有限公司 | Method, apparatus, equipment and medium for entity associated |
CN110162638A (en) * | 2019-04-12 | 2019-08-23 | 淮阴工学院 | A kind of expert's combination proposed algorithm based on figure vector |
CN110245204A (en) * | 2019-06-12 | 2019-09-17 | 桂林电子科技大学 | A kind of intelligent recommendation method based on positioning and knowledge mapping |
Cited By (12)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113495963A (en) * | 2020-03-19 | 2021-10-12 | 复旦大学 | Embedded representation method and device of network security knowledge graph |
CN113495963B (en) * | 2020-03-19 | 2023-03-14 | 复旦大学 | Embedded representation method and device of network security knowledge graph |
CN111708881A (en) * | 2020-05-22 | 2020-09-25 | 国网天津市电力公司 | Text representation learning method introducing incidence relation |
CN112256801A (en) * | 2020-10-10 | 2021-01-22 | 深圳力维智联技术有限公司 | Method, system and storage medium for extracting key entities in entity relationship graph |
CN112256801B (en) * | 2020-10-10 | 2024-04-09 | 深圳力维智联技术有限公司 | Method, system and storage medium for extracting key entity in entity relation diagram |
CN112182249A (en) * | 2020-10-23 | 2021-01-05 | 四川大学 | Automatic classification method and device for aviation safety report |
CN112463989A (en) * | 2020-12-11 | 2021-03-09 | 交控科技股份有限公司 | Knowledge graph-based information acquisition method and system |
CN114283947A (en) * | 2021-12-27 | 2022-04-05 | 北京和兴创联健康科技有限公司 | Health management method and system suitable for surgical patients |
US20230306321A1 (en) * | 2022-03-24 | 2023-09-28 | Chengdu Qinchuan Iot Technology Co., Ltd. | Systems and methods for managing public place in smart city |
CN115617694A (en) * | 2022-11-30 | 2023-01-17 | 中南大学 | Software defect prediction method, system, device and medium based on information fusion |
CN116524723A (en) * | 2023-06-27 | 2023-08-01 | 天津大学 | Truck track anomaly identification method and system |
CN116524723B (en) * | 2023-06-27 | 2023-09-12 | 天津大学 | Truck track anomaly identification method and system |
Also Published As
Publication number | Publication date |
---|---|
CN110704636B (en) | 2021-01-05 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110704636B (en) | Improved Node2 vec-based knowledge graph vector representation method | |
CN109241313B (en) | Image retrieval method based on high-order deep hash learning | |
US7756799B2 (en) | Feature selection based on partial ordered set of classifiers | |
CN110347847B (en) | Knowledge graph complementing method based on neural network | |
CN113326731A (en) | Cross-domain pedestrian re-identification algorithm based on momentum network guidance | |
Chen et al. | An effective feature selection scheme for healthcare data classification using binary particle swarm optimization | |
CN112131402A (en) | PPI knowledge graph representation learning method based on protein family clustering | |
CN114626598A (en) | Multi-modal trajectory prediction method based on semantic environment modeling | |
Rasekh et al. | EDNC: Evolving differentiable neural computers | |
CN116010621B (en) | Rule-guided self-adaptive path generation method | |
CN113836174B (en) | Asynchronous SQL (structured query language) connection query optimization method based on reinforcement learning DQN (direct-to-inverse) algorithm | |
CN116561302A (en) | Fault diagnosis method, device and storage medium based on mixed knowledge graph reasoning | |
US11461656B2 (en) | Genetic programming for partial layers of a deep learning model | |
KR102497362B1 (en) | System for multi-layered knowledge base and processing method thereof | |
Hu et al. | PWSNAS: Powering weight sharing NAS with general search space shrinking framework | |
CN114662693A (en) | Reinforced learning knowledge graph reasoning method based on action sampling | |
CN116975311B (en) | Agricultural pest knowledge graph optimization method, system and computer | |
WO2021059527A1 (en) | Learning device, learning method, and recording medium | |
Wu et al. | How to apply deep reinforcement learning to multi-dimensional intelligent reasoning | |
Chen et al. | Hierarchical multi-label classification incorporating prior information for gene function prediction | |
CN117097624B (en) | Network topology structure enhancement method and device, electronic equipment and storage medium | |
Li et al. | A hybrid reasoning method of knowledge graph for on-line arts education based on reinforcement learning | |
CN116842199B (en) | Knowledge graph completion method based on multi-granularity hierarchy and dynamic embedding | |
Mäntykivi | Genetic Operators in Chain-Structured Neural Architecture Search | |
Mu et al. | Auto-CASH: Autonomous Classification Algorithm Selection with Deep Q-Network |
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 | ||
GR01 | Patent grant | ||
GR01 | Patent grant |