TOPIC MODELING BASED ON ATTRIBUTED GRAPH

News data analysis plays an important role in the development, early warning and management of events. People often use topic modeling for analysis, which helps the government managers to have a macro control of the development of events. However, the existing technologies often obtain topics through LDA-like models or document clustering, and neither of these two methods can study the higher-order semantic relationships among a group of topic-related keywords. Few studies start from the perspective of graph, but they are only based on the simple graph structure. In this paper, we propose to construct a keyword-keyword network based on the structure of attributed graph by using word embedding as attribute, in which the keywords with close edges are regarded as topics. Compared with existing methods, Attributed Graph Topic Model (AGTM) aims to capture deeper topic semantics by introducing word embedding as attributes, capturing topics better than simple graph structures. A large number of experiments show that we achieve good results compared with the baseline model.