Paper Recommendation Based On Random Walk And Multi-features Fusion

In recent years,the rapid economic development has promoted the leap-forward progress of scientific research.Large quantities of scientific and technological achievements of scientific researchers have sprung up.The number of papers stored in various electronic paper databases is almost increasing exponentially.Such rich scientific research achievements have laid a solid foundation for the researches of scientific researchers,but they also cost researchers a lot of time and energy to find relevant papers.If researchers can obtain many related papers by reading only a few academic papers,the efficiency of researchers’ searching papers will be significantly improved.In order to achieve this goal,this paper proposes a paper recommendation algorithm based on random walk and multi-features fusion,which only requires researchers to read a few papers they are interested in,and then recommends series of relevant papers for scientific researchers to read and study,so as to shorten the time for scientific researchers to search for papers and help them put more time and energy into paper reading,studying and scientific research innovations.The paper recommendation algorithm based on random walk and multi-features fusion consists of two parts.In the first part,a network representation learning framework based on random walk is constructed by using only abstract information and citation network structure information,which is an unsupervised paper recommendation model.In this model.Firstly,LDA topic model is carried out based on abstracts of paper nodes in citation network to obtain topic labels of paper nodes,and then network representation learning based on random walk according to topic labels is carried out to obtain vector representation of nodes.Finally,the cosine similarity between node vectors is used as the similarity of papers,and papers with high similarity are recommended.Experiments on real citation networks show that the algorithm has better recommendation effect than Deepwalk and Node2 vec algorithms,and MAP index,NDCG index,precision index,recall index and F1 index have all been improved to some extent.However,it is not enough to use literature abstracts information and network structure information.In order to further improve the effect of recommendation algorithm,this paper proposes a second-step recommendation framework.Based on the research of the first part,this part extracts various features existing between papers,such as content features,topology features and attribute features between citing papers and cited papers,and then the algorithm performs a recommendation algorithm integrating multiple features to recommend relevant papers.The algorithm is actually a classification algorithm.Given two paper nodes,if the classification algorithm judges that the two nodes will be connected,the algorithm will recommend this paper to another connected paper.The classification algorithm uses logistic regression.The experimental results of real citation network show that the algorithm greatly improves the recommendation effect on the basis of the first part of the research.MAP index increases by about 12%,NDCG index increases by about 6%,precision index increases by about 24%,recall index increases by about 17%,F1 index increases by about 30%.Finally,Extra Trees Classifier is used to rank the feature importance of the above three categories of features,and it is not only confirmed that the first part of the research plays a key role in the recommendation effect and found that network topology features and content features play important roles in the identification of relevant papers.