A study on the modeling of topic flows by authors for analyzing personalized trends of research

Abstract

Topic modeling has been one of the powerful unsupervised methods to analyze concepts hidden in a set of documents. The objective of this thesis is to generate author’s topic flows over years. To generate au-thor’s topic flow, we analyzed two models; Author-Topic Model (ATM) and Temporal Author-Topic Model (TAT). ATM finds author’s interest to the topics, but cannot generate topic flow because ATM doesn’t have time concept. To deal with problem, TAT has been proposed. TAT generates author’s topic flow, but indirect topic flow. Indirect topic flow means that author has only author-topic distribution which is the interest to the topics, but not the topic-flow itself. To generate topic flow in TAT, we have to multiply author-topic distribu-tion and topic-year distribution. Therefore, author’s topic flow affected by the topic-year distribution. Because of this reason, Topic flow can’t reflect author’s interest properly. To deal with this problem, we proposed a new topic model called Author-Topic Flow Model (ATFM) which generates direct topic flow. To generate direct topic flow, we extended author node A to have year distribution over all topics. With the proposed model, we performed experiment to prove effectiveness. The result showed that direct topic flow reflects author’s interest better than TAT.