A Study Of Network Feature Representation Using Node2Vec Model

Using complex network to abstract and represent complex systems in the real word,and converting tedious and variety of professional issues into the study of relationship between nodes and nodes,edges and edges has become a mature academic approach applied in ranged studying fields recently.As we digging deeper in the study of complex network,Word2Vec model,which derived from Natural Language Processing field has coincidentally combined with complex network has drawn great attention of researchers.This specific method process and simplify network data and take full utilize of neural networks model and deep learning methods,represent network nodes in the form of low dimension space vectors,which perfectly preserves the network structure as well as relationships between different nodes.Since this method reflects the weak link in the network effectively and retained the long tail effect as well as its high efficiency,rapid and general applicability characteristics when it comes to big data networks,it has been put into use in network clustering,link prediction and visualization and proved efficiently and stability.Nowadays,feature representation learning such as Node2Vec model has been put into use in Biology,Sociology,Economics and performed well.It draws another vision on learning complex network structure and laws,and we can easily reveal the background and theory based on the feature representation learning results behind the blind spots comparing to traditional research methods.On the basis of the studies mentioned above,the author attempts to take a glimpse of the collaboration network of molecular machine research field which was the winner of the Nobel Prize in chemistry as an example of complex network.The author collected academic papers in recent 10 years,clean up and analysis the data and use Node2Vec model to learn the feature representation of collaboration network,perform network clustering,similarity ranking,link prediction and visualization task on the network.In this paper,the author concluded that network feature learning works well on large data complex networks,overcomes the size limits in bibliometric and social network research methods.Network feature learning performs well on clustering analysis,similarity ranking,neighborhood node prediction and visualization tasks,has greater efficiency and stronger adaptability comparing to traditional method.However,link prediction result is far from satisfaction and requires further studies.In conclusion,network feature learning is an effective complement to traditional network research method which pursue for improvement and upgrading in existing research methods,deal solved problems more easily,rather than discover new knowledge and has high potential and remains further study.