This paper comes up with a fast-training model based on random walk to address the problems of slow training speed for representation learning of large-scale cybersecurity knowledge graph and lack of relational representation of head and tail entities,. The model first performs an initial training representation of the entities of the overall knowledge graph by random walk under relational paths, then, a subject-object embedding is designed to learn the syntactical meaning of the relations in the knowledge graph by combining the relation-specific subject embedding with the relation-specific object embedding. Finally, fast training of the knowledge graph is again assisted by random wandering under relational paths. In this paper, extensive experiments are conducted on several datasets and the results are compared with those using several existing models. The results show that the model proposed in this paper can shorten the training time by 1/3 and improve representation by about 3%, effectively improving the representation learning effect while speeding up the training speed of knowledge graph representation learning.

knowledge graph  /  knowledge graph embedding  /  representation learning  /  random walk