Intelligent travel guide systems have grown increasingly popular in recent years. They also benefit a lot from the development of social media, resulting in a large amount of attractions uploaded by users. To tackle this, attractions should be real time classified by user-generated photos automatically to gain better user experience. However, in practice, the given label of photos and text ratings may be incomplete or missing. Moreover, recently, domain adaptation has been applied to deal with few labeled data. Thus, in this paper, we propose a novel framework for automatically attraction classification in leveraging web-harvesting data from search engine and the photos of attractions uploaded by users. Specifically, we assume that top-k web-harvesting images from search engines have correct labels. The classification problem is formulated as a regularized domain adaptation approach. Experiments conducted on the collected real-world data set demonstrated that the promising performance is gained over state-of-the art classification methods.