Image Topic Modeling With Users And Geographic Information

With the rapid development of the Internet, the amount of documents available has reached a very high altitude. Thanks to the popularity of smart phones, more and more documents have got GPS annotation. Traditional topic modeling algorithms treat these GPS-associated documents as normal documents and discard the GPS information. By analyzing the geo-graphical distribution of these documents we can gain useful information such as cultural differences around the world or popularity of specific products in different regions. This calls for effective approaches to study the GPS-associated documents on the Web.This thesis studies the problem of image topic modeling with users and geographic information. By treating tags as words and user as author, an image with GPS information can be viewed as a GPS-associated document. In this paper we introduce a newer and better approach which called Author-related Geographical Topic Modeling (to be called AGTM for short). The method introduces author and GPS information into the topic modeling procedure and uses the information to improve the results. The AGTM algorithm based on the following assumption:given two documents, if the two authors of them are friends or the two GPS locations are not far away from each other, we suppose that these two documents are likely to belong to the same topic.We use Flickr images as dataset and the results shows that AGTM algorithm achieves a better performance for topic modeling. This shows that the assumption mentioned before is reasonable.Moreover, we use VisualRank algorithm to select canonical images for each topic.