Random-walk Based Knowledge Representation Learning

Nowadays,Internet service has become an indispensable part of human life.In recent years,with the rapid development of big data,cloud computing and artificial intelligence technologies,Internet services have become more and more intelligent.Many Internet products,such as search engines and recommendation systems,can return personalized information to different users through big data and artificial intelligence technologies.To accomplish this goal,the machine needs to understand human intentions.Knowledge graph technology can help machines understand human beings.It is a special social network.Nodes in the network represent entities,and edges in the network represent relationships.We can get relationships between entities or judge entities related to them based on entities and relationships.In recent years,deep learning technology has a great success in academia and industry,and has been widely used in tasks such as image recognition,speech recognition and natural language processing,and has made great success.The most important reason that deep learning can succeed is that it can represent pixels or words as low-dimensional vectors.It's called representation vectors.This vector has many practical properties that can help us do a lot of work.The knowledge representation learning is to convert the entities and relationships in the knowledge graph into representation vectors,and finish tasks such as entity prediction and relationship prediction by representing vectors.However,the traditional knowledge representation learning model still has many shortcomings.For example,it does not consider the context in the graph,so the semantics of the representation vector is not sufficient.This paper proposes a KG2 Vec model,which needs to reconstruct the knowledge graph first,transform the entity and relationship into nodes format,and select the node sequence in the knowledge graph through the novel random walk algorithm,and then use Word2 vec model to train the representation vectors of entities and relationships;at the same time,this paper proposes two models,which are called KG2Vec-CBOW and KG2Vec-Skipgram and an optimized training algorithm for training.This optimized training algorithm solves the power-law distribution in knowledge graph and dynamically customizes the random walk parameters of each node.At the end of the paper,we design some related experiments.Since the representation vector should have the same features as the word vector,but the similarity between the entities and the relationships is difficult to define,this paper first defines the concept of similar entities and similar relationships in the knowledge graph,and the recall of similar entities and relationships in the knowledge graph is used as an indicator to measure the performance of the model.Then,the representation vectors of entities and relationships are trained on the FB15 K and WN18 datasets respectively.This paper first tunes the parameters of the random walk model to get the best performance,and use it to train the representation vector.The experimental results show that the representation vector trained by KG2 Vec model has more completed semantics than the vector trained by TransE.