With the purpose of discovering the geographical topics from GPS-assoeiated images, a novel geographical topic model, which is called Author-related Geographical Topic Model (AGTM), is proposed and implemented. AGTM increases the accuracy of geographical topic by introducing author information and their relationships into the traditional modeling process. It assumes that if two authors are friends their documents will have a greater probability of being clustered into the same topic. Under this assumption, AGTM divided authors into several social groups, while social groups are assigned to certain geographic regions under some probability distributions. Experiments performed on a dataset of Flickr images with GPS coordinates and other attributes indicate that our approach achieves a better accuracy and interpretability for modeling geographical topic.%为了对带有地理信息的位置相关图像进行有效的主题建模,提出一种结合作者与地理信息的主题建模算法——AGTM.首先在地理相关主题建模过程中引入数据发布者的信息,并利用发布者之间所存在的好友关系来提高主题建模结果的准确度；在建模过程中基于如果2个图像的发布者相同或者发布者之间存在着好友关系,那么他们所发布的图像数据极可能从属于同一地理相关主题这样一个假设,假设图像发布者以某个概率分布于若干社会群组中,而群组又以一定概率分布于某些地理区域之上,使得在主题模型生成过程中充分利用了以往被忽略的发布者属性.使用Flickr中的图像作为数据集进行实验,结果表明,AGTM算法显著提高了模型的准确度和可解释性.
展开▼
