Social Media Based Travel Data Mining And Personalized Recommendation

Tourism has been speedily developed in the 21 st century, and has been benefitted a lot from Internet. The travel websites carry extensive information that help tourists to make their plans before travelling. Voluminous travel experiences are shared through kinds of social media across the world in every second. Such a tremendous amount of Internet data has potentials to further boost the development of tourism, and to stimulate the research of intelligent tourism. However, it is necessary but difficult to exploit the useful information from the daily-increasing data. Travel information in social media is diverse, redundant,heterogeneous, and discoverable. These properties bring us four challenging problems: 1)it is difficult to obtain the visualized travel information; 2) it is difficult to analyze the heterogeneous travel information; 3) it is hard to retrieval multimedia entity based on text; and4) the personalized travel recommendation model is very hard to build. To address these problems, this thesis makes the following attempts:1. An image reranking method is proposed based on mixed visual feature and graph model, which can collect high-quality images from Internet quickly, to provide guarantees for datum foundation based on travel image analysis in social media. Visual information is the semantic bridge between texts and images. While a single visual feature is insufficient to express the semantics, the semantic latent analysis is employed to learn the mixed feature from multi-features. The learnt mixed visual feature can suppress the redundant information among the multi-features but preserve the latent links between the features. To overcome the disadvantages of classification-based and clustering-based reranking, graph-based reranking method is utilized based on the mixed feature. By using the proposed method, the gathered images can be fed into the travel recommendation systems for travel image analysis and mining.2. A video archaeology method is developed on the basis of user-generated manipulation detection, which can rapidly identify the special video version and eliminate redundant videos from a set of near-duplicated videos, to contribute to travel video analysis in social media. The manipulations of videos are easily generated and uploaded by Internet users.Therefore, manipulation-detectors are necessary for detecting the manipulations, and more importantly, the manipulated relationship between a pair of videos should be accurately accessed. In this method, the video migration map is built based on the relationships of video pairs. This map is then analyzed to explore the migration of near-duplicate videos. This method can be used for the effective video retrieval, the redundant video elimination and the analysis of popular video evolutionary. While videos are regarded as an important visualized information, the high-quality and un-manipulated videos from social media could be a key resource in travel recommendation systems.3. A landmark reranking method is presented based on the fusion of travel heterogeneous information, which can achieve query-dependent landmark retrieval and landmark reranking, to help for travel decision. While most existing travel search engine cannot help users to get appropriate landmark information from massive amounts of travel information,query-dependent landmark search is introduced for the generation of the initial ranking results. The initial landmark retrieval results may exhibit un-related top-k recommendations.To improve this situation, the heterogeneous information of landmarks is employed for reranking. In reranking, the travel heterogeneous information in social media, including images and texts, are mined to extract the multi-latent topical features. Then, the popularity and satisfaction of landmarks, which are calculated by digital information, are introduced for the refinement of landmark reranking. The query-enabled landmark reranking method based on the heterogeneous travel information fusion scheme can facilitate the development of travel intelligent systems and enhance user experiences. Moreover, this method can help users to gain landmark information actively, which can be regarded as a preparing work for the personalized travel recommendation.4. A personalized travel landmark recommendation scheme is designed based on collective intelligence, which can not only make the best of shared user experiences in social media, but also utilize user interactions for personalized landmark recommendation. It is common for users to upload travel information and share experiences through social media. These data contain the collective intelligence, which can be mined to structuralize the reasons of making travel decisions by users. While the travel information has the characteristics of sparsity and diversity, content-based recommendation based on user explicit interaction is employed for landmark recommendation. Used for the knowledge from collective intelligence, manifold regularized classification is introduced to establish the personalized attraction similarity model, which can fuse the heterogeneous information with weight adaption and obtain candidate recommended landmarks. Because user decision is influenced by context, candidate landmarks are ranked by the context of user location, and finally, personalized recommended landmarks are obtained. This proposed method only needs simple interaction of users, and uses collective intelligence properly for personalized landmark recommendation.In conclusion, this thesis aims to innovate the Internet-based travel applications based on the foundation of data mining and machine learning. Specifically, a study of travel information mining and recommendation in social media is conducted. It mainly covers the research topics of image reranking, video archaeology, landmark search and reranking, and personalized landmark recommendation. The possible solutions are also proposed for the problems of 1) the effective acquisition of high-quality visualized information from social media, and 2) the landmark reranking and recommendation based on travel heterogeneous information. Expectedly, the innovative ideas presented in this thesis can solve the problems in natural science research and meet the needs of the developing tourism. Moreover, it will establish the theoretical foundation to accelerate the growth of smart tourism.