Scientific Paper Recommendation Systems Based On Citation Network

The successful applications of recommender systems in social network ande-commerce promote the research upsurge of recommender systems, So that lots ofrecommender algorithms have been proposed. However, the research in scientific paperrecommendation is so rare. The problem of information explosive in scientific paperalso needs to be resolved.The thesis based on citation web, deeply researches the applications of existingrecommender systems algorithms in scientific paper recommendation, and concludesthe advantage and disadvantage of them. After exploring the corresponding improvingmethods, the thesis promotes several new algorithms, and proves their effectiveness andfeasibility. At last, these algorithms successfully improve the recommender quality. Themain details are as follows:1. The thesis analyses the disadvantages of classic collaborative filtering systemsused in scientific paper recommendation. And then, the thesis improves thisrecommender systems to weighted citation web recommender systems, to distinguishthe important papers and unimportant papers, so that, the recommendation results andaccuracy can be promoted. At last, the thesis uses two experiences to test theseimprovements.2. The thesis analyses the problems of the application of link prediction similarityindices in citation web in detail, and discusses the influence on similarity resulted fromoutdegree of resources node, indegree of targets node, co-citation relations andco-reference relations. After that, the thesis proposes similarity indices based on citationweb, and proves the effectiveness of these new indices in link prediction by using threeexperiences.3. The thesis uses the classic mass diffusion and heat conduction algorithms torecommend scientific paper, and analyses it in brief. Then, we analyze and improve theresource allocation problem of mass diffusion. After that, the thesis proposescollaborative similarity between collaborators of scientific papers. By using theimproved resource allocation, the thesis uses PageRank values to distinguish papers with different importance. On the problem of collaborative similarity, we first combineit with mass diffusion in positive way to improve the accuracy. Then, we combine itwith mass diffusion in the opposite way to promote the diversity. These methods makemass diffusion achieve the goals of promoting accuracy and diversity respectively. Thatis, by using different methods to improve mass diffusion, the system can achieve theaccuracy and diversity goals respectively. At last, we use two experiments to provethese improvements.