Difference and connection between graph cube and knowledge map

① Both of them are graphs composed of nodes and edges. But all the entities in the graph network exist objectively, which is a presentation of the real world relationship; Knowledge map mainly presents the potential knowledge structure of the objective world, and entities can be abstract nouns.

② Both of them are heterogeneous information networks, but their tasks are different. KG is a heterogeneous information network with rich knowledge. It pays more attention to expressing relationships and nodes through modeling, and the focus of model learning is the relationship between nodes, so as to better store, extract and reason knowledge. NG modeling task pays more attention to the representation of nodes, and the focus of model learning is the structure of graph network to achieve the purpose of node classification, clustering and link prediction.

Third, graph network representation learning (graph embedding) VS knowledge graph representation learning (knowledge graph embedding)

It can also be called graph embedding learning, which is divided into graph network embedding and knowledge graph embedding. From the origin point of view, the most popular methods of these two tasks, DeepWalk and Transition, are both inspired by word2vec, but the former is inspired by word2vec's processing of text sequences and the prediction of the first word's context. The latter is inspired by word2vec's automatic discovery of implicit relationship (namely king-man = queen-woman).

The similarity between them is the same goal, and both aim to establish a distributed representation of the research object. The difference is that knowledge representation focuses on how to deal with the explicit relationship between entities; Network representation focuses on how to fully consider the complex structural information of nodes in the network (such as communities, etc.). ).

1) Different learning goals

Network representation pays more attention to preserving the topological structure information of the network in the embedded space. Knowledge map indicates that on the basis of retaining structural information, it also pays attention to the importance of relationship and its end-to-end relationship. Knowledge map representation learning is more inclined to relationship modeling, emphasizing the relationship and head-tail relationship on the basis of retaining structural information, and emphasizing the representation of nodes and relationships, which are equally important. Therefore, knowledge map representation learning often indicates relationships, such as the relationship between fruits and kiwifruit.

2) Different learning methods

Network representation learning usually includes three types: models based on matrix decomposition, such as SVD；; Models based on random walking, such as DeepWalk；; Models based on deep neural networks, including CNN and RNN. In addition, there are differences between homogeneous networks and heterogeneous networks, and there are also differences between attribute networks and networks that integrate accompanying information.

Different from this, the typical knowledge map representation algorithms include trans series algorithms, such as TransE, TransR, TransH and so on. A that describes the vector representation of entities and relationships.

Grab the first praise comment, share the report and put it away.

Recommend for you:

Special suggestion

What are some ways to maximize work efficiency?

What are the possible serious consequences of being infected by mites?

Does COVID-19 need asymptomatic treatment?

What are the "saving money" tips in life?