Image-regulated graph topic model for cross-media topic detection
Zhiyi Wang, Liang Li, Chunjie Zhang, Qingming Huang · 2015
In recent years, pictures and videos have become ubiquitous on the Internet, which encourage the development of algorithm that analyze their semantic contents for detecting topics. Among them, topic modeling plays an essential role in discovering topics from document collections. However, with rich auxiliary information (such as geo-information, user-annotated tags, pictures and videos) rising up around the text, traditional topic models show their limitations to discover latent topics effectively from the cross-media data. To address this problem, we propose a novel Image-regulated Graph Topic Model (IGTM), which combines cross-media data together in the modeling process. By utilizing auxiliary relation information among images, IGTM can achieve higher quality underlying topics as image relationships could serve as weakly-supervised information for topic modeling. Experimental results over two cross-media datasets demonstrate the effectiveness of our model.