This paper is motivated by the two-stage combination of probabilistic graphic model and deep learning. We found the DL-based extracted features are conducive for Bayesian learning in probabilistic graphic models. We further develop a tourist recommendation as its application in this field.
A topic model is designed for general representation to describe the characteristics of one given region. Given the collection of photos from different regions, we apply the features extracted from GoogLeNet in representation learning. Comparing the learned representation, the topic model produces a more powerful representation of the regional characteristics.
Geographical Latent Attribute Model (GLAM) suppose that a given geographical region consists of different “attributes” (e.g., infrastructures, attractions, events and activities) and “attributes” are interpreted by different image “clusters”.
When given a photo or a collection of photos, computing the characteristic representation of this collection and the similarity between characteristic representation can achieve the tourist recommendation and the location retrieval.